This is the multi-million dollar .unwrap() story. In a critical path of infrastructure serving a significant chunk of the internet, calling .unwrap() on a Result means you're saying "this can never fail, and if it does, crash the thread immediately."The Rust compiler forced them to acknowledge this could fail (that's what Result is for), but they explicitly chose to panic instead of handle it gracefully. This is textbook "parse, don't validate" anti-pattern.
I know, this is "Monday morning quarterbacking", but that's what you get for an outage this big that had me tied up for half a day.
I’ve led multiple incident responses at a FAANG, here’s my take. The fundamental problem here is not Rust or the coding error. The problem is:
1. Their bot management system is designed to push a configuration out to their entire network rapidly. This is necessary so they can rapidly respond to attacks, but it creates risk as compared to systems that roll out changes gradually.
2. Despite the elevated risk of system wide rapid config propagation, it took them 2 hours to identify the config as the proximate cause, and another hour to roll it back.
SOP for stuff breaking is you roll back to a known good state. If you roll out gradually and your canaries break, you have a clear signal to roll back. Here was a special case where they needed their system to rapidly propagate changes everywhere, which is a huge risk, but didn’t quite have the visibility and rapid rollback capability in place to match that risk.
While it’s certainly useful to examine the root cause in the code, you’re never going to have defect free code. Reliability isn’t just about avoiding bugs. It’s about understanding how to give yourself clear visibility into the relationship between changes and behavior and the rollback capability to quickly revert to a known good state.
Cloudflare has done an amazing job with availability for many years and their Rust code now powers 20% of internet traffic. Truly a great team.
The "coding error" is a somewhat deliberate choice to fail eagerly that is usually safe but doesn't align with the need to do something (propagation of the configuration file) without failing.
I'm sure that there are misapplied guidelines to do that instead of being nice to incoming bot management configuration files, and someone might have been scolded (or worse) for proposing or attempting to handle them more safely.
Does their ring based rollout really truly have to be 0->100% in a few seconds?
I don’t really buy this requirement. At least make it configurable with a more reasonable default for “routine” changes. E.g. ramping to 100% over 1 hour.
As long as that ramp rate is configurable, you can retain the ability to respond fast to attacks by setting the ramp time to a few seconds if you truly think it’s needed in that moment.
Maybe, but in that case maybe have some special casing logic to detect that yes indeed we're under a massive DDOS at this very moment, do a rapid rollout of this thing that will mitigate said DDOS. Otherwise use the default slower one?
Of course, this is all so easy to say after the fact..
Exactly the right take. Even when you want to have rapid changes on your infra, do it at least by region. You can start with the region where the least amount of users are impacted and if everything is fine, there is no elevated number of crashes for example, you can move forward. It was a standard practice at $RANDOM_FAANG when we had such deployments.
In a productive way, this view also shifts the focus to improving the system (visibility etc), empowering the team, rather than focusing on the code which broke (probably strikes fear in the individuals, to do anything!)
Partial disagree. There should be lints against 'unwrap's. An 'expect' at least forces you to write down why you are so certain it can't fail. An unwrap is not just hubris, it's also laziness, and has no place in sensitive code.
And yes, there is a lint you can use against slicing ('indexing_slicing') and it's absolutely wild that it's not on by default in clippy.
> Their bot management system is designed to push a configuration out to their entire network rapidly.
Once every 5m is not "rapidly". It isn't uncommon for configuration systems to do it every few seconds [0].
> While it’s certainly useful to examine the root cause in the code.
Believe the issue is as much an output from a periodic run (clickhouse query) caused by (on the surface, an unrelated change) causing this failure. That is, the system that validated the configuration (FL2) was different to the one that generated it (ML Bot Management DB).
Ideally, it is the system that vends a complex configuration that also vends & tests the library to consume it, or the system that consumes it, does so as if it was "tasting" the configuration first before devouring it unconditionally [1].
Of course, as with all distributed system failures, this is all easier said and done in hindsight.
Isn't rapidly more of how long it takes to get from A to Z rather than how often it is performed? You can push out a configuration update every fortnight but if it goes through all of your global servers in three seconds, I'd call it quite rapid.
It seems people have a blind spot for unwrap, perhaps because it's so often used in example code. In production code an unwrap or expect should be reviewed exactly like a panic.
It's not necessarily invalid to use unwrap in production code if you would just call panic anyway. But just like every unsafe block needs a SAFETY comment, every unwrap in production code needs an INFALLIBILITY comment. clippy::unwrap_used can enforce this.
Yes, I always thought it was wrong to use unwrap in examples. I know, people want to keep examples simple, but it trains developers to use unwrap() as they see that everywhere.
Yes, there are places where it's ok as that blog post explains so well: https://burntsushi.net/unwrap/
But most devs IMHO don't have the time to make the call correctly most of the time... so it's just better to do something better, like handle the error and try to recover, or if impossible, at least do `expect("damn it, how did this happen")`.
Dunno, I think the alternatives have their own pretty significant downsides. All would require front loading more in-depth understanding of error handling and some would just be quite a bit more verbose.
IMO making unwrap a clippy lint (or perhaps a warning) would be a decent start. Or maybe renaming unwrap.
Alternatively you can look at actually innovative programming languages to peek at the next 20 years of innovation.
I am not sure that watching the trendy forefront successfully reach the 1990s and discuss how unwrapping Option is potentially dangerous really warm my heart. I can’t wait for the complete meltdown when they discover effect systems in 2040.
To be more serious, this kind of incident is yet another reminder that software development remains miles away from proper engineering and even key providers like Cloudfare utterly fail at proper risk management.
Celebrating because there is now one popular language using static analysis for memory safety feels to me like being happy we now teach people to swim before a transatlantic boat crossing while we refuse to actually install life boats.
To me the situation has barely changed. The industry has been refusing to put in place strong reliability practices for decades, keeps significantly under investing in tools mitigating errors outside of a few fields where safety was already taken seriously before software was a thing and keeps hiding behind the excuse that we need to move fast and safety is too complex and costly while regulation remains extremely lenient.
I mean this Cloudfare outage probably cost millions of dollars of damage in aggregate between lost revenue and lost productivity. How much of that will they actually have to pay?
Let's try to make effect systems happen quicker than that.
> I mean this Cloudfare outage probably cost millions of dollars of damage in aggregate between lost revenue and lost productivity. How much of that will they actually have to pay?
Probably nothing, because most paying customers of cloudflare are probably signing away their rights to sue Cloudflare for damages by being down for a while when they purchase Cloudflare's services (maybe some customers have SLAs with monetary values attached, I dunno). I honestly have a hard time suggesting that those customers are individually wrong to do so - Cloudflare isn't down that often, and whatever amount it cost any individual customer by being down today might be more than offset by the DDOS protection they're buying.
Anyway if you want Cloudflare regulated to prevent this, name the specific regulations you want to see. Should it be illegal under US law to use `unwrap` in Rust code? Should it be illegal for any single internet services company to have more than X number of customers? A lot of the internet also breaks when AWS goes down because many people like to use AWS, so maybe they should be included in this regulatory framework too.
Yes? Funnily enough, I don't often use indexed access in Rust. Either I'm looping over elements of a data structure (in which case I use iterators), or I'm using an untrusted index value (in which case I explicitly handle the error case). In the rare case where I'm using an index value that I can guarantee is never invalid (e.g. graph traversal where the indices are never exposed outside the scope of the traversal), then I create a safe wrapper around the unsafe access and document the invariant.
If that's the case then hats off. What you're describing is definitely not what I've seen in practice. In fact, I don't think I've ever seen a crate or production codebase that documents infallibility of every single slice access. Even security-critical cryptography crates that passed audits don't do that. Personally, I found it quite hard to avoid indexing for graph-heavy code, so I'm always on the lookout for interesting ways to enforce access safety. If you have some code to share that would be very interesting.
> I don't think I've ever seen a crate or production codebase that documents infallibility of every single slice access.
The smoltcp crate typically uses runtime checks to ensure slice accesses made by the library do not cause a panic. It's not exactly equivalent to GP's assertion, since it doesn't cover "every single slice access", but it at least covers slice accesses triggered by the library's public API. (i.e. none of the public API functions should cause a panic, assuming that the runtime validation after the most recent mutation succeeds).
My rule of thumb is that unchecked access is okay in scenarios where both the array/map and the indices/keys are private implementation details of a function or struct, since an invariant is easy to manually verify when it is tightly scoped as such. I've seen it used it in:
* Graph/tree traversal functions that take a visitor function as a parameter
Sure, these days I'm mostly working on a few compilers. Let's say I want to make a fixed-size SSA IR. Each instruction has an opcode and two operands (which are essentially pointers to other instructions). The IR is populated in one phase, and then lowered in the next. During lowering I run a few peephole and code motion optimizations on the IR, and then do regalloc + asm codegen. During that pass the IR is mutated and indices are invalidated/updated. The important thing is that this phase is extremely performance-critical.
Usually you'd want to write almost all your slice or other container iterations with iterators, in a functional style.
For the 5% of cases that are too complex for standard iterators? I never bother justifying why my indexes are correct, but I don't see why not.
You very rarely need SAFETY comments in Rust because almost all the code you write is safe in the first place. The language also gives you the tool to avoid manual iteration (not just for safety, but because it lets the compiler eliminate bounds checks), so it would actually be quite viable to write these comments, since you only need them when you're doing something unusual.
I didn't restate the context from the code we're discussing: it must not panic. If you don't care if the code panics, then go ahead and unwrap/expect/index, because that conforms to your chosen error handling scheme. This is fine for lots of things like CLI tools or isolated subprocesses, and makes review a lot easier.
So: first, identify code that cannot be allowed to panic. Within that code, yes, in the rare case that you use [i], you need to at least try to justify why you think it'll be in bounds. But it would be better not to.
There are a couple of attempts at getting the compiler to prove that code can't panic (e.g., the no-panic crate).
What about memory allocation - how will you stop that from panicking ? `Vec::resize` will always panic in Rust. And this is just one example out of thousands in the Rust stdlib.
Unless the language addresses no-panic in its governing design or allows try-catch, not sure how you go about this.
That is slowly being addressed, but meanwhile it’s likely you have a reliable upper bound on how much heap your service needs, so it’s a much smaller worry. There are also techniques like up-front or static allocation if you want to make more certain.
Indexing is comparatively rare given the existence of iterators, IMO. If your goal is to avoid any potential for panicking, I think you'd have a harder time with arithmetic overflow.
Your pair of posts is very interesting to me. Can you share with me: What is your programming environment such that you are "fine with allocation failures"? I'm not doubting you, but for me, if I am doing systems programming with C or C++, my program is doomed if a malloc fails! When I saw your post, I immediately thought: Am I doing it wrong? If I get a NULL back from malloc(), I just terminate with an error message.
I mean, yeah, if I am using a library, as an user of this library, I would like to be able to handle the error myself. Having the library decide to panic, for example, is the opposite of it.
Something that allows me to tag annotate a function (or my whole crate) as "no panic", and get a compile error if the function or anything it calls has a reachable panic.
This will allow it to work with many unmodified crates, as long as constant propagation can prove that any panics are unreachable. This approach will also allow crates to provide panicking and non panicking versions of their API (which many already do).
I think the most common solution at the moment is dtolnay's no_panic [0]. That has a bunch of caveats, though, and the ergonomics leave something to be desired, so a first-party solution would probably be preferable.
Yes, I want that. I also want to be able to (1) statically apply a badge on every crate that makes and meets these guarantees (including transitively with that crate's own dependencies) so I can search crates.io for stronger guarantees and (2) annotate my Cargo.toml to not import crates that violate this, so time isn't wasted compiling - we know it'll fail in advance.
On the subject of this, I want more ability to filter out crates in our Cargo.toml. Such as a max dependency depth. Or a frozen set of dependencies that is guaranteed not to change so audits are easier. (Obviously we could vendor the code in and be in charge of our own destiny, but this feels like something we can let crate authors police.)
For iteration, yes. But there's other cases, like any time you have to deal with lots of linked data structures. If you need high performance, chances are that you'll have to use an index+arena strategy. They're also common in mathematical codebases.
It's the same blind spot people have to Java's checked exceptions. People commonly resort to Pokemon exception handling and either blindly ignoring or rethrowing as a runtime exception. When Rust got popular, I was a bit confused by people talking about how great Result it's essentially a checked exception without a stack trace.
"Checked Exceptions Are Actually Good" gang, rise up! :p
I think adoption would have played out very different if there had only been some more syntactic-sugar. For example, an easy syntax for saying: "In this method, any (checked) DeepException e that bubbles up should immediately be replaced by a new (checked) MylayerException(e) that contains the original one as a cause.
We might still get lazy programmers making systems where every damn thing goes into a generic MylayerException, but that mess would still be way easier to fix later than a hundred scattered RuntimeExceptions.
There is also the problem that they decided to make all references nullable, so `NullPointerException`s could appear everywhere. This "forced" them to introduce the escape hatch of `RuntimeException`, which of course was way overused immediately, normalizing it.
I'm with you! Checked exceptions are actually good and the hate for them is super short sighted. The exact same criticisms levied at checked exceptions apply to static typing in general, but people acknowledge the great value static types have for preventing errors at compile time. Checked exceptions have that same value, but are dunked on for some reason.
1. in most cases they don't want to handle `InterruptedException` or `IOException` and yet need to bubble them up. In that case the code is very verbose.
2. it makes lambdas and functions incompatible. So eg: if you're passing a function to forEach, you're forced to wrap it in runtime exception.
3. Due to (1) and (2), most people become lazy and do `throws Exception` which negates most advantages of having exceptions in the first place.
In line-of-business apps (where Java is used the most), an uncaught exception is not a big deal. It will bubble up and gets handled somewhere far up the stack (eg: the server logger) without disrupting other parts of the application. This reduces the utility of having every function throw InterruptedException / IOException when those hardly ever happen.
Yeah, in both cases it's a layering situation, where it's the duty of your code to decide what layers of abstraction need to be be bridged, and to execute on that decision. Translating/wrapping exception-types from deeper functions is the same as translating/wrapping return-types the same places.
I think it comes down to a psychological or use-case issue: People hate thinking about errors and handling them, because it's that hard stuff that always consumes more time than we'd like to think. Not just digitally, but in physical machines too. It's also easier to put off "for later."
Checked exceptions in theory were good, but Java simply did not add facilities to handle or support them well in many APIs. Even the new API's in Java - Streams, etc do not support checked exceptions.
> it's essentially a checked exception without a stack trace
In theory, theory and practice are the same. In practice...
You can't throw a checked exception in a stream, this fact actually underlines the key difference between an exception and a Result: Result is in return position and exceptions are a sort of side effect that has its own control flow. Because of that, once your method throws an Exception or you are writing code in a try block that catches an exception, you become blind to further exceptions of that type, even if you might be able to or required to fix those errors. Results are required to be handled individually and you get syntactic sugar to easily back propagate.
It is trivial to include a stack trace, but stack traces are really only useful for identifying where something occurred, and generally what is superior is attaching context as you back propagate which trivially occurs with judicious use of custom error types with From impls. Doing this means that the error message uniquely defines the origin and paths it passed through without intermediate unimportant stack noise. With exceptions you would always need to catch each exception and rethrow a new exception containing the old to add contextual information, then to avoid catching to much you need variables that will be initialized inside the try block defined outside of the try block. So stack traces are basically only useful when you are doing Pokemon exception handling.
It's a lot lighter: a stack trace takes a lot of overhead to generate; a result has no overhead for a failure. The overhead (panic) only comes once the failure can't be handled. (Most books on Java/C# don't explain that throwing exceptions has high performance overhead.)
Exceptions force a panic on all errors, which is why they're supposed to be used in "exceptional" situations. To avoid exceptions when an error is expected, (eof, broken socket, file not found,) you either have to use an unnatural return type or accept the performance penalty of the panic that happens when you "throw."
In Rust, the stack trace happens at panic (unwrap), which is when the error isn't handled. IE, it's not when the file isn't found, it's when the error isn't handled.
Exceptions do not force panic at all. In most practical situations, an exception unhandled close to where it was thrown will eventually get logged. It's kind of a "local" panic, if you will, that will terminate the specific function, but the rest of the program will remain unaffected. For example, a web server might throw an exception while processing a specific HTTP request, but other HTTP requests are unaffected.
Throwing an exception does not necessarily mean that your program is suddenly in an unsupported state, and therefore does not require terminating the entire program.
checked exceptions failed because when used properly they fossilize method signatures. they're fine if your code will never be changed and they're fine when you control 100% of users of the throwing code. if you're distributing a library... no bueno.
> When Rust got popular, I was a bit confused by people talking about how great Result it's essentially a checked exception without a stack trace.
It's not a checked exception without a stack trace.
Rust doesn't have Java's checked or unchecked exception semantics at the moment. Panics are more like Java's Errors (e.g. OOM error). Results are just error codes on steroids.
That's kind of what I'm saying with the blind spot comment. The words "unwrap" and "expect" should be just as much a scary red flag as the word "panic", but for some reason it seems a lot of people don't see them that way.
Even in lowly Java, they later added to Optional the orElseThrow() method since the name of the get() method did not connote the impact of unwrapping an empty Optional.
Than they are going to write None | Err => yolo() that has the same impact. It is not the syntax or the semantic meaning is the problem here but the fact that there is no monitoring around the elevated error counts after a deployment.
Software engineers tend to get stuck in software problems and thinking that everything should be fixed in code. In reality there are many things outside of the code that you can do to operate unreliable components safely.
The `unsafe` keyword means something specific in Rust, and panicking isn't unsafe by Rust's definition. Sometimes avoiding partial functions just isn't feasible, and an unwrap (or whatever you want to call the method) is a way of providing a (runtime-checked) proof to the compiler that the function is actually total.
unwrap() should effectively work as a Result<> where the user must manually invoke a panic in the failure branch. Make special syntax if a match and panic is too much boilerplate.
This is like an implicit null pointer exception that cannot be statically guarded against.
I want a way to statically block any crates doing this from my dependency chain.
Not sure what you're saying with the "work as a Result<>" part...unwrap is a method on Result. I think you're just saying the unwrap/expect methods should be eliminated?
> In production code an unwrap or expect should be reviewed exactly like a panic.
An unwrap should never make it to production IMHO. It's fine while prototyping, but once the project gets closer to production it's necessary to just grep `uncheck` in your code and replace those that can happen with a proper error management and replace those that cannot happen with `expect`, with a clear justification of why they cannot happen unless there's a bug somewhere else.
I would say, sure, if you feel the same way about panic calls making to production. In other words, review all of them the same way. Because writing unwrap/expect is exactly the same as writing “if error, panic”.
Isn't the point of this article that pieces of infrastructure don't go down to root causes, but due to bad combinations of components that are correct individually? After reading "engineering a safer world", I find root cause analysis rather reductionistic, because it wasn't just an unwrap, it was that the payload was larger than normal, because of a query that didn't select by database, because a clickhouse made more databases visible. Hard to say "it was just due to an unwrap" imo. Especially in terms of how to fix an issue going forwards. I think the article lists a lot of good ideas, that aren't just "don't unwrap", like enabling more global kill switches for features, or eliminating the ability for core dumps or other error reports to overwhelm system resources.
You're right. A good postmortem/root cause analysis would START from "unwrap" and continue from there.
You might start with a basic timeline of what happened, then you'd start exploring: why did this change affect so many customers (this would be a line of questioning to find a potential root cause), why did it take so long to discover or recover (this might be multiple lines of questioning), etc.
> This is the multi-million dollar .unwrap() story.
That's too semantic IMHO. The failure mode was "enforced invariant stopped being true". If they'd written explicit code to fail the request when that happened, the end result would have been exactly the same.
> If the `.unwrap()` was replaced with `.expect("Feature config is too large!")` it would certainly make the outage shorter.
It wouldn't, not meaningfully. The outage was caused by change in how they processed the queries. They had no way to observe the changes, nor canaries to see that change is killing them. Plus, they would still need to manually feed and restart services that ingested bad configs.
`expect` would shave a few minutes; you would still spend hours figuring out and fixing it.
Granted, using expect is better, but it's not a silver bullet.
In general for unexpected errors like these the internal function should log the error, and I assume it was either logged or they can quickly deduce reason based on the line number.
>> This is the multi-million dollar .unwrap() story.
> That's too semantic IMHO. The failure mode was "enforced invariant stopped being true". If they'd written explicit code to fail the request when that happened, the end result would have been exactly the same.
Problem is, the enclosing function (`fetch_features`) returns a `Result`, so the `unwrap` on line #82 only serves as a shortcut a developer took due to assuming `features.append_with_names` would never fail. Instead, the routine likely should have worked within `Result`.
> Instead, the routine likely should have worked within `Result`.
But it's a fatal error. It doesn't matter whether it's implicit or explicit, the result is the same.
Maybe you're saying "it's better to be explicit", as a broad generalization I don't disagree with that.
But that has nothing to do with the actual bug here, which was that the invariant failed. How they choose to implement checking and failing the invariant in the semantics of the chosen language is irrelevant.
> Today, many friends pinged me saying Cloudflare was down. As a core developer of the first generation of Cloudflare FL, I'd like to share some thoughts.
> This wasn't an attack, but a classic chain reaction triggered by “hidden assumptions + configuration chains” — permission changes exposed underlying tables, doubling the number of lines in the generated feature file. This exceeded FL2's memory preset, ultimately pushing the core proxy into panic.
> Rust mitigates certain errors, but the complexity in boundary layers, data flows, and configuration pipelines remains beyond the language's scope. The real challenge lies in designing robust system contracts, isolation layers, and fail-safe mechanisms.
> Hats off to Cloudflare's engineers—those on the front lines putting out fires bear the brunt of such incidents.
> Technical details: Even handling the unwrap correctly, an OOM would still occur. The primary issue was the lack of contract validation in feature ingest. The configuration system requires “bad → reject, keep last-known-good” logic.
> Why did it persist so long? The global kill switch was inadequate, preventing rapid circuit-breaking. Early suspicion of an attack also caused delays.
> Why not roll back software versions or restart?
> Rollback isn't feasible because this isn't a code issue—it's a continuously propagating bad configuration. Without version control or a kill switch, restarting would only cause all nodes to load the bad config faster and accelerate crashes.
> Why not roll back the configuration?
> Configuration lacks versioning and functions more like a continuously updated feed. As long as the ClickHouse pipeline remains active, manually rolling back would result in new corrupted files being regenerated within minutes, overwriting any fixes.
This tweet thread invokes genuine despair in me. Do we really have to outsource even our tweets to LLMs? Really? I mean, I get spambots and the like tweeting mass-produced slop. But what compels a former engineer of the company in question to offer LLM-generated "insight" to the outage? Why? For what purpose?
* For clarity, I am aware that the original tweets are written in Chinese, and they still have the stench of LLM writing all over them; it's not just the translation provided in the above comment.
To be fair, this failed in the non-rust path too because the bot management returned that all traffic was a bot. But yes, FL2 needs to catch panics from individual components but I’m not sure if failing open is necessarily that much better (it was in this case but the next incident could easily be the result of failing open).
But more generally you could catch the panic at the FL2 layer to make that decision intentional - missing logic at that layer IMHO.
Catching panic probably isn’t a great idea if there’s any unsafe code in the system. (Do the unsafe blocks really maintain heap invariants if across panics?)
Unsafe blocks have nothing to do with it. Yes - they maintain all the same invariants as safe blocks or those unsafe blocks are unsound regardless of panics. But there’s millions of way to architect this (eg a supervisor process that notices which layer in FL2 is crashing and just completely disables that layer when it starts up the proxy again. There’s challenges here because then you have to figure out what constitutes a perma crashing (eg what if it’s just 20% of all sites? Do you disable?). And in the general case you have the fail open/fail close decision anyway which you should just annotate individual layers with.
But the bigger change is to make sure that config changes roll out gradually instead of all at once. That’s the source of 99% of all widespread outages
I think the parent is implying that the panic should be "caught" via a supervisor process, Erlang-style, rather than implying the literal use of `catch_unwind` to resume within the same process.
Supervisor is the brutalist way. But catch_unwind may be needed for perf and other reasons.
But ultimately it’s not the panic that’s the problem but a failure to specify how panics within FL2 layers should be handled; each layer is at least one team and FL2’s job is providing a safe playground for everyone to safely coexist regardless of the misbehavior of any single component
But as always such failures are emblematic of multiple things going wrong at once. You probably want to end up using both catch_unwind for the typical case and the supervisor for the case where there’s a segfault in some unsafe code you call or native library you invoke.
I also mention the fundamental tension of do you want to fail open or closed. Most layers should probably fail open. Some layers (eg auth) it’s safer to fail closed.
This is assuming that the process could have done anything sensible while it had the malformed feature file. It might be in this case that this was one configuration file of several and maybe the program could have been built to run with some defaults when it finds this specific configuration invalid, but in the general case, if a program expects a configuration file and can't do anything without it, panicking is a normal thing to do. There's no graceful handling (beyond a nice error message) a program like Nginx could do on a syntax error in its config.
The real issue is further up the chain where the malformed feature file got created and deployed without better checks.
I do not think that if the bot detection model inside your big web proxy has a configuration error it should panic and kill the entire proxy and take 20% of the internet with it. This is a system that should fail gracefully and it didn't.
> The real issue
Are there single "real issues" with systems this large? There are issues being created constantly (say, unwraps where there shouldn't be, assumptions about the consumers of the database schema) that only become apparent when they line up.
I don't know too much about how the feature file distribution works but in the event of failure to read a new file, wouldn't logging the failure and sticking with the previous version of the file be preferable?
That's exactly the point (ie just prior to distribution) where a simple sanity check should have been run and the config replacement/update pipeline stopped on failure. When they introduced the 200 entry limit memory optimised feature loader it should have been a no-brainer to insert that sanity check in the config production pipeline.
Yea, Rust is safe but it’s not magic. However Nginx doesn’t panic on malformed config. It exits with hopefully a helpful error code and message. The question is then could the cloudflare code have exited cleanly in a way that made recovery easier instead of just straight panicking.
Would expect with a message meet that criteria of exiting with a more helpful error message? From the postmortem it seems to me like they just didn’t know it even was panicing
One feature failing like this should probably log the error and fail closed. It shouldn't take down everything else in your big proxy that sits in front of your entire business.
I'm not a fan of rust, but I don't think that is the only takeaway. All systems have assumptions about their input and if the assumption is violated, it has to be caught somewhere. It seems like it was caught too deep in the system.
Maybe the validation code should've handled the larger size, but also the db query produced something invalid. That shouldn't have ever happened in the first place.
> It seems like it was caught too deep in the system.
Agreed, that's also my takeaway.
I don't see the problem being "lazy programmers shouldn't have called .unwrap()". That's reductive. This is a complex system and complex system failures aren't monocausal.
The function in question could have returned a smarter error rather than panicking, but what then? An invariant was violated, and maybe this system, at this layer, isn't equipped to take any reasonable action in response to that invariant violation and dying _is_ the correct thing to do.
But maybe it could take smarter action. Maybe it could be restarted into a known good state. Maybe this service could be supervised by another system that would have propagated its failure back to the source of the problem, alerting operators that a file was being generated in such a way that violated consumer invariants. Basically, I'm describing a more Erlang model of failure.
Regardless, a system like this should be able to tolerate (or at least correctly propagate) a panic in response to an invariant violation.
Safe things should be easy, dangerous things should be hard.
This .unwrap() sounds too easy for what it does, certainly much easier than having an entire try..catch block with an explicit panic. Full disclosure: I don't actually know Rust.
Any project has to reason about what sort of errors can be tolerated gracefully and which cannot. Unwrap is reasonable in scenarios you expect to never be reached, because otherwise your code will be full of all sorts of possible permutations and paths that are harder to reason about and may cascade into extremely nuanced or subtle errors.
Rust also has a version of unwrap called "expect" where you provide a string that logs why the unwrap occurred. It's similar, but for pieces of code that are crucial it could be a good idea to require all 'unwraps' to instead be 'expects' so that people at least are forced to write down a reason why they believe the unwrap can never be reached.
Not panicking code is tedious to write. It is not realistic to expect everything to be non panic. There is a reason that panicking exists in the first place.
Them calling unwrap on a limit check is the real issue imo. Everything that takes in external input should assume it is bad input and should be fuzz tested imo.
In the end, what is the point of having a limit check if you are just unwrapping on it
Using the question mark operator [1] and even adding in some anyhow::context goes a long way to being able to fail fast and return an Err rather then panicking.
Sure you need to handle Results all the way up the stack but at least it forces you to think about how those nested parts of your app will fail as you travel back up the stack.
Swift has implicit unwrap (!), and explicit unwrap (?).
I don't like to use implicit unwrap. Even things that are guaranteed to be there, I treat as explicit (For example, (self.view?.isEnabled ?? false), in a view controller, instead of self.view.isEnabled).
So what happens if it ends up being nil? How does your app react?
In this particular case, I would rather crash. It’s easier to spot in a crash report and you get a nice stack trace.
Silent failure is ultimately terrible for users.
Note: for the things I control I try to very explicitly model state in such a way as I never need to force unwrap at all. But for things beyond my control like this situation, I would rather end the program than continue with a state of the world I don’t understand.
Yeah @IBOutlets are generally the one thing that are allowed to be implicitly-unwrapped optionals. They go along with using storyboards & xibs files with Interface Builder. I agree that you really should just crash if you are attempting to access one and it is nil. Either you have done something completely incorrect with regards to initializing and accessing parts of your UI and want to catch that in development, or something has gone horribly, horribly, horribly with UIKit/AppKit and storyboard/xib files are not being loaded properly by the system.
A good tool for catching stuff during development, is the humble assert()[0]. We can use precondition()[1], to do the same thing, in ship code.
The main thing is, is to remain in control, as much as possible. Rather than let the PC leave the stack frame, throw the error immediately when it happens.
> Silent failure is ultimately terrible for users.
Agreed.
Unfortunately, crashes in iOS are “silent failures,” and are a loss of control.
What this practice does, is give me the option to handle the failure “noisily,” and in a controlled manner; even if just emitting a log entry, before calling a system failure. That can be quite helpful, in threading. Also, it gives me the option to have a valid value applied, if there’s a structural failure.
But the main reason that I do that with @IBOutlets, is that it forces me to acknowledge, throughout the rest of the code, that it’s an optional. I could always treat implicit optionals as if they were explicit, anyway. This just forces me to.
I have a bunch of practices that folks can laugh at, but my stuff works pretty effectively, and I sleep well.
That's such a bad take after reading the article. If you're going to write a system that preallocates and is based on hard assumptions about max size - the panic/unwrap approach is reasonable.
The config bug reaching prod without this being caught and pinpointed immediately is the strange part.
It's reasonable when testing protocols exercise the panic scenario. This is the problem with punting on error recovery. Nobody checks faults that propagate across domains of responsibility.
This is a bummer. The unwrap()'ing function already returned a result and should have just propagated the error. Presumably the caller could have handled more sensibly than just panic'ing.
Some languages and style guides simply forbid throwing exceptions without catching / proper recovery. Google C++ bans exceptions and the main mechanism for propogating errors is `absl::Status` which the caller has to check. Not familiar with Rust but it seems unwrap is such a thing that would be banned.
> Not familiar with Rust but it seems unwrap is such a thing that would be banned.
Panics aren't exceptions, any "panic" in Rust can be thought of as an abort of the process (Rust binaries have the explicit option to implement panics as aborts). Companies like Dropbox do exactly this in their similar Rust-based systems, so it wouldn't surprise me if Cloudflare does the same.
"Banning exceptions" wouldn't have done anything here, what you're looking for is "banning partial functions" (in the Haskell sense).
Yeah I know but isn't unwrap basically a trivial way to: (1) give up catching the exception/error (the E part in `Result<T, E>`) that the callee throws; and also (2) escalate it to the point that nothing can catch it. It has such a benign name.
Unwrap is used in places where in C++ you would just have undefined behavior. It wouldn't make any more sense to blanket ban it than it would ban ever dereferencing a pointer just in case its null - even if you just checked that it wasn't null.
Rust's foo: Option<&T> is rust's rough equivalent to C++'s const T* foo. The C++ *foo is equivalent to the rust unsafe{ *foo.unwrap_unchecked() }, or in safe code *foo.unwrap() (which changes the undefined behavior to a panic).
Rust's unwrap isn't the same as std::expected::value. The former panics - i.e. either aborts the program or unwinds depending on context and is generally not meant to be handled. The latter just throws an exception that is generally expected to be handled. Panics and exceptions use similar machinery (at least they can depending on compiler options) but they are not equivalent - for example nested panics in destructors always abort the program.
In code that isn't meant to crash `unwind` should be treated as a sign saying that "I'm promising that this will never happen", but just like in C++ where you promise that pointers you deference are valid and signed integers you add don't overflow making promises like that is a necessary part of productive programming.
Linting is not good enough. The compiler should refuse to compile the code without it marked with an explicit annotation. Too much Rust code is panic happy since using casual use of `unwrap` is perma-etched into everyone's minds by the amount of demo code out there using unwrap.
> same like chess, engine is better than human grandmaster because its solvable math field
Might be worth noting that your description of chess is slightly incorrect. Chess technically isn't solved in the sense that the optimal move is known for any arbitrary position is known; it's just that chess engines are using what amounts to a fancy brute force for most of the game and the combination of hardware and search algorithm produces a better result than the human brain does. As such, chess engines are still capable of making mistakes, even if actually exploiting them is a challenge.
> because these thing called BEST MOVE and BAD MOVE there in chess
The thing is that there is no known general objective criteria for "best" and "bad" moves. The best we have so far is based on engine evaluations, but as I said before that is because chess engines are better at searching the board's state space than humans, not because chess engines have solved chess in the mathematical sense. Engines are quite capable of misevaluating positions, as demonstrated quite well by the Top Chess Engine Championship [0] where one engine thinks it made a good move while the other thinks that move is bad, and this is especially the case when resources are limited.
The closest we are to solving chess are via tablebases, which are far from covering the entire state space and are basically as much of an exemplar of pure brute force as you can get.
> "chess engines are still capable of making mistakes", I'm sorry no
If you think chess engines are infalliable, then why does the Top Chess Engine Championship exist? Surely if chess engines could not make mistakes they would always agree on a position's evaluation and what move should be made, and therefore such an exercise would be pointless?
> inaccurate yes but not mistake
From the perspective to attaining perfect play an inaccuracy is a mistake.
"The thing is that there is no known general objective criteria for "best" and "bad" moves."
are you playing chess or not?????? if you playing chess then its oblivious how to differentiate bad move and best move
Yes it is objective, these thing called best move not without reason
"If you think chess engines are infalliable, then why does the Top Chess Engine Championship exist?"
to create better chess engine like what do even talking about here????, are you saying just because there are older bad engine that mean this thing is pointless ????
if you playing chess up to a decent level 1700+ (like me), you know that these argument its wrong and I assure you to learn chess to a decent level
up until that point that you know high level chess is brute force games and therefore solvable math
> if you playing chess then its oblivious how to differentiate bad move and best move
The key words in what I said are "general" and "objective". Yes, it's possible to determine "good" or "bad" moves in specific positions. There's no known method to determine "good" or "bad" moves in arbitrary positions, as would be required for chess to be considered strongly solved.
Furthermore, if it's "obvious" how to differentiate good and bad moves then we should never see engines blundering, right?
So (for example) how do you explain this game between Stockfish and Leela where Stockfish blunders a seemingly winning position [0]? After 37... Rdd8 both Stockfish and Leela think white is clearly winning (Stockfish's evaluation is +4.00, while Leela's evaluation is +3.81), but after 38. Nxb5 Leela's evaluation plummets to +0.34 while Stockfish's evaluation remains at +4.00. In the end, it turns out Leela was correct after 40... Rxc6 Stockfish's evaluation also drops from +4.28 to 0.00 as it realizes that Leela has a forced stalemate.
Or this game also between Stockfish and Leela where Leela blunders into a forced mating sequence and doesn't even realize it for a few moves [1]?
Engines will presumably always play what they think is the "best" move, but clearly sometimes this "best" move is wrong. Evidently, this means differentiating "good" and "bad" moves is not always obvious.
> Yes it is objective, these thing called best move not without reason
If it's objective, then why is it possible for engines to disagree on whether a move is good or bad, as they do in the above example and others?
> to create better chess engine like what do even talking about here????
The ability to create better chess engines necessarily implies that chess engines can and do make mistakes, contrary to what you asserted.
> are you saying just because there are older bad engine that mean this thing is pointless ????
No. What I'm saying is that your explanation for why chess engines are better than humans is wrong. Chess engines are not better than humans because they have solved chess in the mathematical sense; chess engines are better than humans because they search the state space faster and more efficiently than humans (at least until you reach 7 pieces on the board).
> up until that point that you know high level chess is brute force games and therefore solvable math
"Solvable" and "solved" are two very different things. Chess is solvable, in theory. Chess is very far from being solved.
I'm not completely sure I agree. I mean, I do agree about the .unwrap() culture being a bug trap. But I don't think this example qualifies.
The root cause here was that a file was mildly corrupt (with duplicate entries, I guess). And there was a validation check elsewhere that said "THIS FILE IS TOO BIG".
But if that's a validation failure, well, failing is correct? What wasn't correct was that the failure reached production. What should have happened is that the validation should have been a unified thing and whatever generated the file should have flagged it before it entered production.
And that's not an issue with function return value API management. The software that should have bailed was somewhere else entirely, and even there an unwrap explosion (in a smoke test or pre-release pass or whatever) would have been fine.
It sounds to me like there was validation, but the system wasn't designed for the validation to ever fail - at which point crashing is the only remaining option. You've essentially turned it into an assertion error rather than a parsing/validation error.
Ideally every validation should have a well-defined failure path. In the case of a config file rotation, validation failure of the new config could mean keeping the old config and logging a high-priority error message. In the case of malformed user-provided data, it might mean dropping the request and maybe logging it for security analysis reasons. In the case of "pi suddenly equals 4" checks the most logical approach might be to intentionally crash, as there's obviously something seriously wrong and application state has corrupted in such a way that any attempt to continue is only going to make things worse.
But in all cases there's a reason behind the post-validation-failure behavior. At a certain point leaving it up to "whatever happens on .unwrap() failure" isn't good enough anymore.
In addition, it looks like this system wasn't on any kind of 1%/10%/50%/100% rollout gating. Such a rollout would trivially have shown the poison input killing tasks.
To me it reads like there was a gradual rollout of the faulty software responsible for generating the config files, but those files are generated on approximately one machine, then propogated across the whole network every 5 minutes.
> Bad data was only generated if the query ran on a part of the cluster which had been updated. As a result, every five minutes there was a chance of either a good or a bad set of configuration files being generated and rapidly propagated across the network.
It looks like changing the permissions triggered creation of a new feature file, and it was ingestion of that file leading to blowing a size limit that crashed the systems.
The file should be versioned and rollout of new versions should be staged.
(There is definitely a trade-off; often times in the security critical path, you want to go as fast as possible because changes may be blocking a malicious actor. But if you move too fast, you break things. Here, they had a potential poison input in the pathway for synchronizing this state and Murphy's Law suggests it was going to break eventually, so the question becomes "How much damage can we tolerate when it does?")
> It looks like changing the permissions triggered creation of a new feature file, and it was ingestion of that file leading to blowing a size limit that crashed the systems.
That feature file is generated every 5 minutes at all times; the change to permissions was rolled out gradually over the clickhouse cluster, and whether a bad version of that file was generated depended on whether the part of the cluster that had the bad permissions generated the file.
> This is textbook "parse, don't validate" anti-pattern.
How so? “Parse, don’t validate” implies converting input into typed values that prevent representation of invalid state. But the parsing still needs to be done correctly. An unchecked unwrap really has nothing to do with this.
At Facebook they name certain "escape hatch" functions in a way that inescapably make them look like a GIANT EYESORE. Stuff like DANGEROUSLY_CAST_THIS_TO_THAT, or INVOKE_SUPER_EXPENSIVE_ACTION_SEE_YOU_ON_CODE_REVIEW. This really drives home the point that such things must not be used except in rare extraordinary cases.
If unwrap() were named UNWRAP_OR_PANIC(), it would be used much less glibly. Even more, I wish there existed a super strict mode when all places that can panic are treated as compile-time errors, except those specifically wrapped in some may_panic_intentionally!() or similar.
React.__SECRET_INTERNALS_DO_NOT_USE_OR_YOU_WILL_BE_FIRED comes to mind. I did have to reach to this before, but it certainly works for keeping this out of example code and other things like reading other implementations without the danger being very apparent.
At some point it was renamed to __CLIENT_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE which is much less fun.
right and if the language designers named it UNWRAP_OR_PANIC() then people would rightfully be asking why on earth we can't just use a try-catch around code and have an easier life
But a panic can be caught and handled safely (e.g. via std:: panic tools). I'd say that this is the correct use case for exceptions (ask Martin Fowler, of all people).
There is already a try/catch around that code, which produces the Result type, which you can presumptuously .unwrap() without checking if it contains an error.
Instead, one should use the question mark operator, that immediately returns the error from the current function if a Result is an error. This is exactly similar to rethrowing an exception, but only requires typing one character, the "?".
How so? An exception is a value that's given the closest, conceptually appropriate, point that was decided to handle the value, allowing you to keep your "happy path" as clean code, and your "exceptional circumstances" path at the level of abstraction that makes sense.
It's way less book-keeping with exceptions, since you, intentionally, don't have to write code for that exceptional behavior, except where it makes sense to. The return by value method, necessarily, implements the same behavior, where handling is bubbled up to the conceptually appropriate place, through returns, but with much more typing involved. Care is required for either, since not properly bubbling up an exception can happen in either case (no re-raise for exceptions, no return after handling for return).
There are many many pages of text discussing this topic, but having programmed in both styles, exceptions make it too easy for programmer to simply ignore them. Errors as values force you to explicitly handle it there, or toss it up the stack. Maybe some other languages have better exception handling but in Python it’s god awful. In big projects you can basically never know when or how something can fail.
I would claim the opposite. If you don't catch an exception, you'll get a halt.
With return values, you can trivially ignore an exception.
let _ = fs::remove_file("file_doesn't_exist");
or
value, error = some_function()
// carry on without doing anything with error
In the wild, I've seen far more ignoring return errors, because of the mechanical burden of having type handling at every function call.
This is backed by decades of writing libraries. I've tried to implement libraries without exceptions, and was my admittedly cargo-cult preference long ago, but ignoring errors was so prevalent among the users of all the libraries that I now always include a "raise" type boolean that defaults to True for any exception that returns an error value, to force exceptions, and their handling, as default behavior.
> In big projects you can basically never know when or how something can fail.
How is this fundamentally different than return value? Looking at a high level function, you can't know how it will fail, you just know it did fail, from the error being bubbled up through the returns. The only difference is the mechanism for bubbling up the error.
Maybe some water is required for this flame war. ;)
I'd categorize them more as "event handlers" than "hidden". You can't know where the execution will go at a lower level, but that's the entire point: you don't care. You put the handlers at the points where you care.
Correction: unchecked exceptions are hidden control flow. Checked exceptions are quite visible, and I think that more languages should use them as a result.
...and you can? try-catch is usually less ergonomic than the various ways you can inspect a Result.
try {
data = some_sketchy_function();
} catch (e) {
handle the error;
}
vs
result = some_sketchy_function();
if let Err(e) = result {
handle the error;
}
Or better yet, compare the problematic cases where the error isn't handled:
data = some_sketchy_function();
vs
data = some_sketchy_function().UNWRAP_OR_PANIC();
In the former (the try-catch version that doesn't try or catch), the lack of handling is silent. It might be fine! You might just depend on your caller using `try`. In the latter, the compiler forces you to use UNWRAP_OR_PANIC (or, in reality, just unwrap) or `data` won't be the expected type and you will quickly get a compile failure.
What I suspect you mean, because it's a better argument, is:
which is fair, although how often is it really the right thing to let all the errors from 4 independent sources flow together and then get picked apart after the fact by inspecting `e`? It's an easier life, but it's also one where subtle problems constantly creep in without the compiler having any visibility into them at all.
Unwrap isn't a synonym for laziness, it's just like an assertion, when you do unwrap() you're saying the Result should NEVER fail, and if does, it should abort the whole process. What was wrong was the developer assumption, not the use of unwrap.
It also makes it very obvious in the code, something very dangerous is happening here. As a code reviewer you should see an unwrap() and have alarm bells going off. While in other languages, critical errors are a lot more hidden.
> What was wrong was the developer assumption, not the use of unwrap.
How many times can you truly prove that an `unwrap()` is correct and that you also need that performance edge?
Ignoring the performance aspect that often comes from a hat-trick, to prove such a thing you need to be wary of the inner workings of a call giving you a `Return`. That knowledge is only valid at the time of writing your `unwrap()`, but won't necessarily hold later.
Also, aren't you implicitly forcing whoever changes the function to check for every smartass dev that decided to `unwrap` at their callsite? That's bonkers.
I doubt that this unwrap was added for performance reasons; I suspect it was rather added because the developer temporarily didn't want to deal with what they thought was an unlikely error case while they were working on something else; and no other system recognized that the unwrap was left in and flagged it before it was deployed on production servers.
If I were Cloudflare I would immediately audit the codebase for all uses of unwrap (or similar rust panic idioms like expect), ensure that they are either removed or clearly documented as to why it's worth crashing the program there, and then add a linter to their CI system that will fire if anyone tries to check in a new commit with unwrap in it.
Panics are for unexpected error conditions, like your caller passed you garbage. Results are for expected errors, like your caller passed you something but it's your job to tell if it's garbage.
So the point of unwrap() is not to prove anything. Like an assertion it indicates a precondition of the function that the implementer cannot uphold. That's not to say unwrap() can't be used incorrectly. Just that it's a valid thing to do in your code.
> No more than returning an int by definition means the method can return -2.
What? Returning an int does in fact mean that the method can return -2. I have no idea what your argument is with this, because you seem to be disagreeing with the person while actually agreeing with them.
The difference is functions which return Result have explicitly chosen to return a Result because they can fail. Sure, it might not fail in the current implementation and/or configuration, but that could change later and you might not know until it causes problems. The type system is there to help you - why ignore it?
Because it would be a huge hassle to go into that library and write an alternate version that doesn't return a Result. So you're stuck with the type system being wrong in some way. You can add error-handling code upfront but it will be dead code at that point in time, which is also not good.
Yeah, I think I expressed wrongly here. A more correct version would be: "when you do unwrap() you're saying that an error on this particular path shouldn't be recoverable and we should fail-safe."
It's a little subtler than this. You want it to be easy to not handle an error while developing, so you can focus on getting the core logic correct before error-handling; but you want it to be hard to deploy or release the software without fully handling these checks. Some kind of debug vs release mode with different lints seems like a reasonable approach.
> This showed up to Internet users trying to access our customers' sites as an error page indicating a failure within Cloudflare's network.
As a visitor to random web pages, I definitely appreciated this—much better than their completely false “checking the security of your connection” message.
> The issue was not caused, directly or indirectly, by a cyber attack or malicious activity of any kind. Instead, it was triggered by a change to one of our database systems' permissions
Also appreciate the honesty here.
> On 18 November 2025 at 11:20 UTC (all times in this blog are UTC), Cloudflare's network began experiencing significant failures to deliver core network traffic. […]
> Core traffic was largely flowing as normal by 14:30. We worked over the next few hours to mitigate increased load on various parts of our network as traffic rushed back online. As of 17:06 all systems at Cloudflare were functioning as normal.
Why did this take so long to resolve? I read through the entire article, and I understand why the outage happened, but when most of the network goes down, why wasn't the first step to revert any recent configuration changes, even ones that seem unrelated to the outage? (Or did I just misread something and this was explained somewhere?)
Of course, the correct solution is always obvious in retrospect, and it's impressive that it only took 7 minutes between the start of the outage and the incident being investigated, but it taking a further 4 hours to resolve the problem and 8 hours total for everything to be back to normal isn't great.
> much better than their completely false “checking the security of your connection” message
The exact wording (which I can easily find, because a good chunk of the internet gives it to me, because I’m on Indian broadband):
> example.com needs to review the security of your connection before proceeding.
It bothers me how this bald-faced lie of a wording has persisted.
(The “Verify you are human by completing the action below.” / “Verify you are human” checkbox is also pretty false, as ticking the box in no way verifies you are human, but that feels slightly less disingenuous.)
Because we initially thought it was an attack. And then when we figured it out we didn’t have a way to insert a good file into the queue. And then we needed to reboot processes on (a lot) of machines worldwide to get them to flush their bad files.
Thanks for the explanation! This definitely reminds me of CrowdStrike outages last year:
- A product depends on frequent configuration updates to defend against attackers.
- A bad data file is pushed into production.
- The system is unable to easily/automatically recover from bad data files.
(The CrowdStrike outages were quite a bit worse though, since it took down the entire computer and remediation required manual intervention on thousands of desktops, whereas parts of Cloudflare were still usable throughout the outage and the issue was 100% resolved in a few hours)
It'd be fun to read more about how you all procedurally respond to this (but maybe this is just a fixation of mine lately). Like are you tabletopping this scenario, are teams building out runbooks for how to quickly resolve this, what's the balancing test for "this needs a functional change to how our distributed systems work" vs. "instead of layering additional complexity on, we should just have a process for quickly and maybe even speculatively restoring this part of the system to a known good state in an outage".
Probably because it was the London team that was actively investigating the incident and initially came to the conclusion that it may be a DDoS while being unable to authenticate to their own systems.
Question from a casual bystander, why not have a virtual/staging mini node that receives these feature file changes first and catches errors to veto full production push?
Or you do have something like this but the specific db permission change in this context only failed in production
I think the reasoning behind this is because of the nature of the file being pushed - from the post mortem:
"This feature file is refreshed every few minutes and published to our entire network and allows us to react to variations in traffic flows across the Internet. It allows us to react to new types of bots and new bot attacks. So it’s critical that it is rolled out frequently and rapidly as bad actors change their tactics quickly."
In this case, the file fails quickly. A pretest that consists of just attempting to load the file would have caught it. Minutes is more than enough time to perform such a check.
Side thought as we're working on 100% onchain systems (for digital assets security, different goals):
Public chains (e.g. EVMs) can be a tamper‑evident gate that only promotes a new config artifact if (a) a delay or multi‑sig review has elapsed, and (b) a succinct proof shows the artifact satisfies safety invariants like ≤200 features, deduped, schema X, etc.
That could have blocked propagation of the oversized file long before it reached the edge :)
Just asking out of curiosity, but roughly how many staff would've been involved in some way in sorting out the issue? Either outside regular hours or redirected from their planned work?
I'm curious about how their internal policies work such that they are allowed to publish a post mortem this quickly, and with this much transparency.
Any other large-ish company, there would be layers of "stakeholders" that will slow this process down. They will almost always never allow code to be published.
Well… we have a culture of transparency we take seriously. I spent 3 years in law school that many times over my career have seemed like wastes but days like today prove useful. I was in the triage video bridge call nearly the whole time. Spent some time after we got things under control talking to customers. Then went home. I’m currently in Lisbon at our EUHQ. I texted John Graham-Cumming, our former CTO and current Board member whose clarity of writing I’ve always admired. He came over. Brought his son (“to show that work isn’t always fun”). Our Chief Legal Officer (Doug) happened to be in town. He came over too. The team had put together a technical doc with all the details. A tick-tock of what had happened and when. I locked myself on a balcony and started writing the intro and conclusion in my trusty BBEdit text editor. John started working on the technical middle. Doug provided edits here and there on places we weren’t clear. At some point John ordered sushi but from a place with limited delivery selection options, and I’m allergic to shellfish, so I ordered a burrito. The team continued to flesh out what happened. As we’d write we’d discover questions: how could a database permission change impact query results? Why were we making a permission change in the first place? We asked in the Google Doc. Answers came back. A few hours ago we declared it done. I read it top-to-bottom out loud for Doug, John, and John’s son. None of us were happy — we were embarrassed by what had happened — but we declared it true and accurate. I sent a draft to Michelle, who’s in SF. The technical teams gave it a once over. Our social media team staged it to our blog. I texted John to see if he wanted to post it to HN. He didn’t reply after a few minutes so I did. That was the process.
> I texted John to see if he wanted to post it to HN. He didn’t reply after a few minutes so I did
Damn corporate karma farming is ruthless, only a couple minute SLA before taking ownership of the karma. I guess I'm not built for this big business SLA.
How do you guys handle redaction? I'm sure even when trusted individuals are in charge of authoring, there's still a potential of accidental leakage which would probably be best mitigated by a team specifically looking for any slip ups.
I mean the CEO posted the post-mortem so there aren't that many layers of stakeholders above. For other post-mortems by engineers, Matthew once said that the engineering team is running the blog and that he wouldn't event know how to veto even if he wanted [0]
The person who posted both this blog article and the hacker news post, is Matthew Prince, one of highly technical billionaire founders of cloudflare. I'm sure if he wants something to happen, it happens.
I’m surprised by how quickly they put together a post mortem, let alone how quickly they were willing to publish it to the world.
Meanwhile, on X, white nationalist supremacists / Groypers / America First America Only types are spreading a conspiracy theory that Cloudflare’s outage was caused by having too many Indians employees and H1Bs. Example:
Why give this sort of content more visibility/reach?
I'm sure that's not your intent, so I hope my comment gives you an opportunity to reflect on the effects of syndicating such stupidity, no matter what platform it comes from.
Mainly to make others aware of what’s happening in the context of this Cloudflare outage. Sure I can avoid giving it visibility/reach but it’s growing and proliferating on its own, and I think ignoring it isn’t going to stop it so I am hoping awareness will help. I’ve noticed a huge rise in open racism against Chinese and Indian and workers of other origin, even when they’re here on a legal visa that we have chosen as a nation to grant for our own benefit.
The legislation that MTG (Marjorie Taylor Green) just proposed a few days ago to ban H1B entirely, and the calls to ban other visa types, is going to have a big negative impact on the tech industry and American innovation in general. The social media stupidity is online but it gives momentum to the actual real life legislation and other actions the administration might take. Many congress people are seeing the online sentiment and changing their positions in response, unfortunately.
The unwrap: not great, but understandable. Better to silently run with a partial config while paging oncall on some other channel, but that's a lot of engineering for a case that apparently is supposed to be "can't happen".
The lack of canary: cause for concern, but I more or less believe Cloudflare when they say this is unavoidable given the use case. Good reason to be extra careful though, which in some ways they weren't.
The slowness to root cause: sheer bad luck, with the status page down and Azure's DDoS yesterday all over the news.
The broken SQL: this is the one that I'd be up in arms about if I worked for Cloudflare. For a system with the power to roll out config to ~all of prod at once while bypassing a lot of the usual change tracking, having this escape testing and review is a major miss.
> That feature file, in turn, doubled in size. The larger-than-expected feature file was then propagated to all the machines that make up our network.
> The software running on these machines to route traffic across our network reads this feature file to keep our Bot Management system up to date with ever changing threats. The software had a limit on the size of the feature file that was below its doubled size. That caused the software to fail.
I'm no FAANG 10x engineer, and I appreciate things can be obvious in hindsight, but I'm somewhat surprised that engineering at the level of Cloudflare does not:
1. Push out files A/B to ensure the old file is not removed.
2. Handle the failure of loading the file (for whatever reason) by automatically reloading the old file instead and logging the error.
It reads a lot like the Crowdstrike SNAFU. Machine-generated configuration file b0rks-up the software that consumes it.
The "...was then propagated to all the machines that make up our network..." followed by "....caused the software to fail." screams for a phased rollout / rollback methodology. I get that "...it’s critical that it is rolled out frequently and rapidly as bad actors change their tactics quickly" but today's outage highlights that rapid deployment isn't all upside.
The remediation section doesn't give me any sense that phased deployment, acceptance testing, and rapid rollback are part of the planned remediation strategy.
At my employer, we have a small script that automatically checks such generated config files. It does a diff between the old and the new version, and if the diff size exceeds a threshold (either total or relative to the size of the old file), it refuses to do the update, and opens a ticket for a human to look over it.
It has somewhat regularly saved us from disaster in the past.
I don't think this system is best thought of as "deployment" in the sense of CI/CD; it's a control channel for a distributed bot detection system that (apparently) happens to be actuated by published config files (it has a consul-template vibe to it, though I don't know if that's what it is).
Code and Config should be treated similarly. If you would use a ring based rollout, canaries, etc for safely changing your code, then any config that can have the same impact must also use safe rollout techniques.
That's why I likened it Crowdstrike. It's a signature database that blew up the consumer of said database. (You probably caught my post mid-edit, too. You may be replying to the snarky paragraph I felt better of and removed.)
Edit: Similar to Crowdstrike, the bot detector should have fallen-back to its last-known-good signature database after panicking, instead of just continuing to panic.
> work has already begun on how we will harden them against failures like this in the future. In particular we are:
> Hardening ingestion of Cloudflare-generated configuration files in the same way we would for user-generated input
> Enabling more global kill switches for features
> Eliminating the ability for core dumps or other error reports to overwhelm system resources
> Reviewing failure modes for error conditions across all core proxy modules
Absent from this list are canary deployments and incremental or wave-based deployment of configuration files (which are often as dangerous as code changes) across fault isolation boundaries -- assuming CloudFlare has such boundaries at all. How are they going to contain the blast radius in the future?
This is something the industry was supposed to learn from the CrowdStrike incident last year, but it's clear that we still have a long way to go.
Also, enabling global anything (i.e., "enabling global kill switches for features") sounds like an incredibly risky idea. One can imagine a bug in a global switch that transforms disabling a feature into disabling an entire system.
They require the bot management config to update and propagate quickly in order to respond to attacks - but this seems like a case where updating a since instance first would have seen the panic and stopped the deploy.
I wonder why clickhouse is used to store the feature flags here, as it has it's own duplication footguns[0] which could have also easily lead to a query blowing up 2/3x in size. oltp/sqlite seems more suited, but i'm sure they have their reasons
I don't think sqlite would come close to their requirements for permissions or resilience, to name a couple. It's not the solution for every database issue.
Also, the link you provided is for eventual deduplication at the storage layer, not deduplication at query time.
I think you're oversimplifying the problem they had, and I would encourage you to dive in to the details in the article. There wasn't a problem with the database, it was with the query used to generate the configs. So if an analogous issue arose with a query against one of many ad-hoc replicated sqlite databases, you'd still have the failure.
I love sqlite for some things, but it's not The One True Database Solution.
Global configuration is useful for low response times to attacks, but you need to have very good ways to know when a global config push is bad and to be able to rollback quickly.
In this case, the older proxy's "fail-closed" categorization of bot activity was obviously better than the "fail-crash", but every global change needs to be carefully validated to have good characteristics here.
Having a mapping of which services are downstream of which other service configs and versions would make detecting global incidents much easier too, by making the causative threads of changes more apparent to the investigators.
It seems they had this continous rollout for the config service, but the services consuming this were affected even by small percentage of these config providers being faulty, since they were auto updating every few minutes their configs. And it seems there is a reason for these updating so fast, presumably having to react to threat actors quickly.
It's in everyone's interest to mitigate threats as quickly as possible. But it's of even greater interest that a core global network infrastructure service provider not DOS a significant proportion of the Internet by propagating a bad configuration too quickly. The key here is to balance responsiveness against safety, and I'm not sure they struck the right balance here. I'm just glad that the impact wasn't as long and as severe as it could have been.
In my 30 years of reliability engineering, I've come to learn that this is a distinction without a difference.
People think of configuration updates (or state updates, call them what you will) as inherently safer than code updates, but history (and today!) demonstrates that they are not. Yet even experienced engineers will allow changes like these into production unattended -- even ones who wouldn't dare let a single line of code go live without being subject to the full CI/CD process.
They narrowed down the actual problem to some Rust code in the Bot Management system that enforced a hard limit on the number of configuration items by returning an error, but the caller was just blindly unwrapping it.
A dormant bug in the code is usually a condition precedent to incidents like these. Later, when a bad input is given, the bug then surfaces. The bug could have laid dormant for years or decades, if it ever surfaced at all.
The point here remains: consider every change to involve risk, and architect defensively.
If they're going to yeet configs into production, they ought to at least have plenty of mitigation mechanisms, including canary deployments and fault isolation boundaries. This was my primary point at the root of this thread.
And I hope fly.io has these mechanisms as well :-)
It's great that you're working on regionalization. Yes, it is hard, but 100x harder if you don't start with cellular design in mind. And as I said in the root of the thread, this is a sign that CloudFlare needs to invest in it just like you have been.
I recoil from that last statement not because I have a rooting interest in Cloudflare but because the last several years of working at Fly.io have drilled Richard Cook's "How Complex Systems Fail"† deep into my brain, and what you said runs aground of Cook #18: Failure free operations require experience with failure.
If the exact same thing happens again at Cloudflare, they'll be fair game. But right now I feel people on this thread are doing exactly, precisely, surgically and specifically the thing Richard Cook and the Cook-ites try to get people not to do, which is to see complex system failures as predictable faults with root causes, rather than as part of the process of creating resilient systems.
Suppose they did have the cellular architecture today, but every other fact was identical. They'd still have suffered the failure! But it would have been contained, and the damage would have been far less.
Fires happen every day. Smoke alarms go off, firefighters get called in, incident response is exercised, and lessons from the situation are learned (with resulting updates to the fire and building codes).
Yet even though this happens, entire cities almost never burn down anymore. And we want to keep it that way.
As Cook points out, "Safety is a characteristic of systems and not of their components."
What variant of cellular architecture are you referring to? Can you give me a link or few? I'm fascinated by it and I've led a team to break up a monolithic solution running on AWS to a cellular architecture. The results were good, but not magic. The process of learning from failures did not stop, but it did change (for the better).
No matter what architecture, processes, software, frameworks, and systems you use, or how exhaustively you plan and test for every failure mode, you cannot 100% predict every scenario and claim "cellular architecture fixes this". This includes making 100% of all failures "contained". Not realistic.
If your AWS service is properly regionalized, that’s the minimum amount of cellular architecture required. Did your service ever fail in multiple regions simultaneously?
Cellular architecture within a region is the next level and is more difficult, but is achievable if you adhere to the same principles that prohibit inter-regional coupling:
It wasn't worth thinking about. I'm not going to defend myself against arguments and absolute claims I didn't make. The key word here is mitigation, not perfection.
> If your AWS service is properly regionalized, that’s the minimum amount of cellular architecture required
Amazon has had multi-region outages due to pushing bad configs, so it’s extremely difficult to believe whatever you are proposing solves that exact problem by relying on multi-regions.
Come to think of it, Cloudflare’s outage today is another good counterexample.
It has been a very, very long time since AWS had a simultaneous failure across multiple regions. Even customers impacted by the loss of Route 53 control plane functionality in last month’s us-east-1 were able to gracefully fail over to a backup region if they configured failover records in advance, had Application Recovery Controller set up, or fronted their APIs or websites with Global Accelerator.
Customers survive incidents on a daily basis by failing over across regions (even in the absence of an AWS regional failure, they can fail due to a bad deployment or other cause). The reason you don’t hear about it is because it works.
Reframe this problem: instead of bot rules being propagated, it's the enrollment of a new customer or a service at an existing customer --- something that must happen at Cloudflare several times a second. Does it still make sense to you to think about that in terms of "pushing new configuration to prod"?
Those aren't the facts before us. Also, CRUD operations relating to a specific customer or user tend not to cause the sort of widespread incidents we saw today.
it's always a config push. people rollout code slowly but don't have the same mechanisms for configs. But configs are code, and this is a blind spot that causes an outsized percentage of these big outages.
absolutely not normal, this is why in my opinion it took them so long to understand the core issue. instead of a nice error message and a backtrace saying something like "failed to parse config feature names" they thought they were under attack because the service was just crashing instead.
"Throwing us off and making us believe this might have been an attack was another apparent symptom we observed: Cloudflare’s status page went down. The status page is hosted completely off Cloudflare’s infrastructure with no dependencies on Cloudflare. While it turned out to be a coincidence, it led some of the team diagnosing the issue to believe that an attacker may be targeting both our systems as well as our status page."
Unfortunately they do not share, what caused the status page to went down as well. (Does this happen often? Otherwise a big coincidence it seems)
The status page is hosted on AWS Cloudfront, right? It sure looks like Cloudfront was overwhelmed by the traffic spike, which is a bit concerning. Hope we'll see a post from their side.
It looks a lot like a CloudFront error we randomly saw today from one of our engineers in South America. I suspect there was a small outage in AWS but can't prove it.
Probably non zero number of companies use cloudfront and other cdns as fallback for cloudflare or running a blended cdn so not surprising to see other cdns hit with a thundering herd when cloudflare went down
it seems like a good chance that despite thinking their status page was completely independent of cloudfront, enough of the internet is dependent on cloudfront now that they're simply wrong about the status page's independence.
Everyone is hating on unwrap, but to me the odd and more interesting part is that it took 3 hours to figure this out? Even with a DDoS red herring, shouldn’t there have been a crash log or telemetry anomaly correlated? Also, shouldn’t the next steps and resolution focus more on this aspect, since it’s a high leverage tool for identifying any outage caused by a panic rather than just preventing a recurrence of random weird edge case #9999999?
Yes this is the weird part for me. With good monitoring, the panic at unwrap should have been detected immediately. I assume they weren't looking at the right place, but still. If you use Sentry for example, a brand new panic should be pretty visible.
I have nowhere near the experience managing such complex systems, but I can empathize with this. In a high-pressure situations the most obvious things get missed. If someone is convinced System X is at fault, your mind can make leaps to justify every other degraded system is a downstream effect of that. Cause and effect can get switched.
Sometimes you have smart people in the room who dig deeper and fish it out, but you cannot always rely on that.
I have plenty of empathy, having been in plenty of similar situations. It's not a matter of "I can't BELIEVE it took that long" (although it is a bit surprising) so much as that I disagree with the key takeaways here in the HN comments section and in the blog itself, which focus strongly on fixing rare edge case issues (the bad ClickHouse query and a bad config file causing a panic via unwrap), rather than reducing MTTR for all issues by improving the debug and monitoring experience.
I'm also suspicious that
> Eliminating the ability for core dumps or other error reports to overwhelm system resources
from the blog had a lot more to do with the issue than perhaps the narrative is letting on.
This took one of the three hours; it seems to have taken from 11:28 to 13:37 to recognize that the configuration file panic was the cause of the issue.
Why does cloudflare allow unwraps in their code? I would've assumed they'd have clippy lints stopping that sort of thing. Why not just match with { ok(value) => {}, Err(error) => {} } the function already has a Result type.
At the bare minimum they could've used an expect("this should never happen, if it does database schema is incorrect").
The whole point of errors as values is preventing this kind of thing.... It wouldn't have stopped the outage but it would've made it easy to diagnose.
If anyone at cloudflare is here please let me in that codebase :)
Not a cloudflare employee but I do write a lot of Rust. The amount of things that can go wrong with any code that needs to make a network call is staggeringly high. unwrap() is normal during development phase but there are a number of times I leave an expect() for production because sometimes there's no way to move forward.
Yeah it seems likely that even if there wasn't an unwrap, there would have been some error handling that wouldn't have panicked the process, but would have still left it inoperable if every request was instead going through an error path.
I'm in a similar boat, at the very leas an expect can give hits to what happened. However this can also be problematic if your a library developer. Sometimes rust is expected to never panic especially in situations like WASM. This is a major problem for companies like Amazon Prime Video since they run in a WASM context for their TV APP. Any panic crashes everything. Personally I usually just either create a custom error type (preferred) or erase it away with Dyn Box Error (no other option). Random unwraps and expects haunt my dreams.
Long time ago Google had a very similar incident where ddos protection system ingested a bad config and took everything down. Except it was auto resolved in like four minutes by an automatic rollback system before oncall was even able to do anything. Perhaps Cloudflare should invest in a system like that
The most surprising thing to me here is that it took 3 hours to root cause, and points to a glaring hole in the platform observability. Even taking into account the fact that the service was failing intermittently at first, it still took 1.5 hours after it started failing consistently to root cause. But the service was crashing on startup. If a core service is throwing a panic at startup like that, it should be raising alerts or at least easily findable via log aggregation. It seems like maybe there was some significant time lost in assuming it was an attack, but it also seems strange to me that nobody was asking "what just changed?", which is usually the first question I ask during an incident.
> thread fl2_worker_thread panicked: called Result::unwrap() on an Err value
I don't use Rust, but a lot of Rust people say if it compiles it runs.
Well Rust won't save you from the usual programming mistake. Not blaming anyone at cloudflare here. I love Cloudflare and the awesome tools they put out.
end of day - let's pick languages | tech because of what we love to do. if you love Rust - pick it all day. I actually wanna try it for industrial robot stuff or small controllers etc.
there's no bad language - just occassional hiccups from us users who use those tools.
You misunderstand what Rust’s guarantees are. Rust has never promised to solve or protect programmers from logical or poor programming. In fact, no such language can do that, not even Haskell.
Unwrapping is a very powerful and important assertion to make in Rust whereby the programmer explicitly states that the value within will not be an error, otherwise panic. This is a contract between the author and the runtime. As you mentioned, this is a human failure, not a language failure.
Pause for a moment and think about what a C++ implementation of a globally distributed network ingress proxy service would look like - and how many memory vulnerabilities there would be… I shudder at the thought… (n.b. nginx)
This is the classic example of when something fails, the failure cause over indexes on - while under indexing on the quadrillions of memory accesses that went off without a single hitch thanks to the borrow checker.
I postulate that whatever the cost in millions or hundreds of millions of dollars by this Cloudflare outage, it has paid for more than by the savings of safe memory access.
Well, no, most Rust programmers misunderstand what the guarantees are because they keep parroting this quote. Obviously the language does not protect you from logic errors, so saying "if it compiles, it works" is disingenuous, when really what they mean is "if it compiles, it's probably free of memory errors".
No, the "if it compiles, it works" is genuinely about the program being correct rather than just free of memory errors, but it's more of a hyperbolic statement than a statement of fact.
It's a common thing I've experienced and seen a lot of others say that the stricter the language is in what it accepts the more likely it is to be correct by the time you get it to run. It's not just a Rust thing (although I think Rust is _stricter_ and therefore this does hold true more of the time), it's something I've also experienced with C++ and Haskell.
So no, it's not a guarantee, but that quote was never about Rust's guarantees.
> Pause for a moment and think about what a C++ implementation of a globally distributed network ingress proxy service would look like - and how many memory vulnerabilities there would be… I shudder at the thought
I mean thats an unfalsifiable statement, not really fair. C is used to successfully launch spaceships.
Whereas we have a real Rust bug that crashed a good portion of the internet for a significant amount of time. If this was a C++ service everyone would be blaming the language, but somehow Rust evangelicals are quick to blame it on "unidiomatic Rust code".
A language that lets this easily happen is a poorly designed language. Saying you need to ban a commonly used method in all production code is broken.
Only formal proof languages are immune to such properties. Therefore all languages are poorly designed by your metric.
Consider that the set of possible failures enabled by language design should be as small as possible.
Rust's set is small enough while also being productive. Until another breakthrough in language design as impactful as the borrow checker is invented, I don't imagine more programmers will be able to write such a large amount of safe code.
> Rust won't save you from the usual programming mistake.
Disagree. Rust is at least giving you an "are you sure?" moment here. Calling unwrap() should be a red flag, something that a code reviewer asks you to explain; you can have a linter forbid it entirely if you like.
No language will prevent you from writing broken code if you're determined to do so, and no language is impossible to write correct code in if you make a superhuman effort. But most of life happens in the middle, and tools like Rust make a huge difference to how often a small mistake snowballs into a big one.
> Disagree. Rust is at least giving you an "are you sure?" moment here. Calling unwrap() should be a red flag, something that a code reviewer asks you to explain; you can have a linter forbid it entirely if you like.
No one treats it like that and nearly every Rust project is filled with unwraps all over the place even in production system like Cloudflare's.
> you can have a linter forbid it entirely if you like.
It would be better if that would be the other way round "linter forbids it unless you ask it not to". Never wrong to allow users to shoot themself in the foot, but it should be explicit.
Yep, unwrap() and unsafe are escape hatches that need very good justifications. It's fine for casual scripts where you don't care if it crashes. For serious production software they should be either banned, or require immense scrutiny.
What people are saying is that idiomatic prod rust doesn't use unwrap/expect (both of which panic on the "exceptional" arm of the value) --- instead you "match" on the value and kick the can up a layer on the call chain.
What happens to it up the callstack? Say they propagated it up the stack with `?`. It has to get handled somewhere. If you don't introduce any logic to handle the duplicate databases, what else are you going to do when the types don't match up besides `unwrap`ing, or maybe emitting a slightly better error message? You could maybe ignore that module's error for that request, but if it was a service more critical than bot mitigation you'd still have the same symptom of getting 500'd.
as they say in the post, these files get generated every 5 minutes and rolled out across their fleet.
so in this case, the thing farther up the callstack is a "watch for updated files and ingest them" component.
that component, when it receives the error, can simply continue using the existing file it loaded 5 minutes earlier.
and then it can increment a Prometheus metric (or similar) representing "count of errors from attempting to load the definition file". that metric should be zero in normal conditions, so it's easy to write an alert rule to notify the appropriate team that the definitions are broken in some way.
that's not a complete solution - in particular it doesn't necessarily solve the problem of needing to scale up the fleet, because freshly-started instances won't have a "previous good" definition file loaded. but it does allow for the existing instances to fail gracefully into a degraded state.
in my experience, on a large enough system, "this could never happen, so if it does it's fine to just crash" is almost always better served by a metric for "count of how many times a thing that could never happen has happened" and a corresponding "that should happen zero times" alert rule.
Given that the bug was elsewhere in the system (the config file parser spuriously failed), it’s hard to justify much of what you suggested.
Panics should be logged, and probably grouped by stack trace for things like prometheus (outside of process). That handles all sorts of panic scenarios, including kernel bugs and hardware errors, which are common at cloudflare scale.
Similarly, mitigating by having rapid restart with backoff outside the process covers far more failure scenarios with far less complexity.
One important scenario your approach misses is “the watch config file endpoint fell over”, which probably would have happened in this outage if 100% of servers went back to watching all of a sudden.
Sure, you could add an error handler for that too, and for prometheus is being slow, and an infinite other things. Or, you could just move process management and reporting out of process.
The way I’ve seen this on a few older systems was that they always keep the previous configuration around so it can switch back. The logic is something like this:
1. At startup, load the last known good config.
2. When signaled, load the new config.
3. When that passes validation, update the last-known-good pointer to the new version.
That way something like this makes the crash recoverable on the theory that stale config is better than the service staying down. One variant also recorded the last tried config version so it wouldn’t even attempt to parse the latest one until it was changed again.
For Cloudflare, it’d be tempting to have step #3 be after 5 minutes or so to catch stuff which crashes soon but not instantly.
The config file subsystem was where the bug lived, not the code with the unwrap, so this sort of change is a special case of “make the unwrap never fail and then fix the API so it is not needed”.
> Well Rust won't save you from the usual programming mistake
This is not a Rust problem. Someone consciously chose to NOT handle an error, possibly thinking "this will never happen". Then someone else conconciouly reviewed (I hope so) a PR with an unwrap() and let it slide.
And people doing testing failed to ignore their excuse of this never happening and still testing it. With this kind of systems you need the separate group that just ignores any "this will never happen" and still checks what happens if it does.
Now it might be that it was tested, but then ignored or deprioritised by management...
> I don't use Rust, but a lot of Rust people say if it compiles it runs.
Do you grok what the issue was with the unwrap, though...?
Idiomatic Rust code does not use that. The fact that it's allowed in a codebase says more about the engineering practices of that particular project/module/whatever. Whoever put the `unwrap` call there had to contend with the notion that it could panic and they still chose to do it.
It's a programmer error, but Rust at least forces you to recognize "okay, I'm going to be an idiot here". There is real value in that.
While I agree that Rust got it right by being more explicit, a lot of bugs in C/C++ can also easily avoided with good engineering practices. The Rust argument that it is mainly the fault of the programming language with C/C++ was always a huge and unfair exaggeration. Now with this entirely predictable ".unwrap" desaster (in general, not necessarily this exact scenarious), the "no true Rustacean would have put unwrap in production" fallacy is sad and funny at the same time.
Unwrap is controversial. The problem is that if you remove it, it makes the bar even higher for newcomers to Rust. One solution is to make it unsafe (along with panic).
We shouldn't be having critical internet-wide outages on a monthly basis. Something is systematically wrong with the way we're architecting our systems.
Cloudflare, Azure, and other single points of failure are solving issues inherent to webhosting, and those problems have become incredibly hard due to the massive scale of bad actors and the massive complexity of managing hardware and software.
What would you propose to fix it? The fixed cost of being DDoS-proof is in the hundreds of millions of dollars.
> Costs to architect systems that serve millions of request daily have gone down. Not up.
I never said serving millions of requests is more expensive. Protecting your servers is more expensive.
> Hell, I would be very curious to know the costs to keep HackerNews running. They probably serve more users than my current client.
HN uses Cloudflare. You're making my point for me. If you included the fixed costs that Cloudflare's CDN/proxy is giving to HN incredibly cheaply, then running HN at the edge with good performance (and protecting it from botnets) would costs hundreds of millions of dollars.
> People want to chase the next big thing to write it on their CV, not architect simple systems that scale. (Do they even need to scale?)
Again, attacking your own straw men here.
Writing high-throughput web applications is easier than ever. Hosting them on the open web is harder than ever.
While I heavily frown upon using `unwrap` and `expect` in Rust code and make sure to have Clippy tell me about every single usage of them, I also understand that without them Rust might have been seen as an academic curiosity language.
They are escape hatches. Without those your language would never take off.
But here's the thing. Escape hatches are like emergency exits. They are not to be used by your team to go to lunch in a nearby restaurant.
---
Cloudflare should likely invest in better linting and CI/CD alerts. Not to mention isolated testing i.e. deploy this change only to a small subset and monitor, and only then do a wider deployment.
Hindsight is 20/20 and we can all be smartasses after the fact of course. But I am really surprised because lately I am only using Rust for hobby projects and even I know I should not use `unwrap` and `expect` beyond the first iteration phases.
---
I have advocated for this before but IMO Rust at this point will benefit greatly from disallowing those unsafe APIs by default in release mode. Though I understand why they don't want to do it -- likely millions of CI/CD pipelines will break overnight. But in the interim, maybe a rustc flag we can put in our `Cargo.toml` that enables such a stricter mode? Or have that flag just remove all the panicky API _at compile time_ though I believe this might be a Gargantuan effort and is likely never happening (sadly).
In any case, I would expect many other failures from Cloudflare but not _this_ one in particular.
Say you panic, now you need to be an external system that catches this panic and report back; and do something meanwhile to recover your system.
If you think about it, it’s not really different from handling the bubbled up error inside of Rust. You don’t (?) your results and your errors go away, they just move up the chain.
I agree the failure is in testing but what you can and should do is raise in alert in your APM system before the runtime panic, in the code path that is deemed impossible to hit.
I am not trashing on them, I've made such mistakes in the past, but I do expect more from them is all.
And you will not believe how many alerts I got for the "impossible" errors.
I do agree there was not too much that could have been done, yes. But they should have invested in more visibility and be more thorough. I mean, hobbyist Rust devs seem to do that better.
It was just a bit disappointing for me. As mentioned above, I'd understand and sympathise with many other mistakes but this one stung a bit.
There's certainly a discipline involved here, but it's usually something like guaranteeing all threads are unwind safe (via AssertUnwindSafe) and logging stack traces when your process keeps dying/can't be started after a fixed number of retries. Which would lead you to the culprit immediately.
I'm just pushing back a bit on the idea that unwrap() is unsafe - it's not, and I wouldn't even call it a foot gun. The code did what it was written to do, when it saw the input was garbage it crashed because it couldn't make sense of what to do next. That's a desirable property in reliable systems (of course monitoring that and testing it is what makes it reliable/fixable in the first place).
We don't disagree, my main point was a bit broader and admittedly hijacked the original topic a bit, namely: `unwrap` and `expect` make many Rust devs too comfortable and these two are very tempting mistresses.
Using those should be done in an extremely disciplined manner. I agree that there are many legitimate uses but in the production Rust code I've seen this has rarely been the case. People just want to move on and then forget to circle back and add proper error handling. But yes, in this case that's not quite true. Still, my point that an APM alert should have been raised on the "impossible" code path before panicking, stands.
Oh for sure. I even think there deserve to be lints like "no code path reachable from main() is unwind-unsafe" which is a heavy hammer for many applications (like one-off CLI utils) but absolutely necessary for something like a long-lived daemon or server that's responsible for critical infrastructure.
Lots of people here are (perhaps rightfully) pointing to the unwrap() call being an issue. That might be true, but to me the fact that a reasonably "clean" panic at a defined line of code was not quickly picked up in any error monitoring system sounds just as important to investigate.
Assuming something similar to Sentry would be in use, it should clearly pick up the many process crashes that start occurring right as the downtime starts. And the well defined clean crashes should in theory also stand out against all the random errors that start occuring all over the system as it begins to go down, precisely because it's always failing at the exact same point.
I integrated Turnstile with a fail-open strategy that proved itself today. Basically, if the Turnstile JS fails to load in the browser (or in a few specific frontend error conditions), we allow the user to submit the web form with a dummy challenge token. On the backend, we process the dummy token like normal, and if there is an error or timeout checking Turnstile's siteverify endpoint, we fail open.
Of course, some users were still blocked, because the Turnstile JS failed to load in their browser but the subsequent siteverify check succeeded on the backend. But overall the fail-open implementation lessened impact to our customers nonetheless.
Fail-open with Turnstile works for us because we have other bot mitigations that are sufficient to fall back on in the event of a Cloudflare outage.
Only if they are able to block the siteverify check performed by our backend server. That's not the kind of attack we are trying to mitigate with Turnstile.
Having the feature table pivoted (with 200 feature1, feature2, etc columns) meant they had to do meta queries to system.columns to get all the feature columns which made the query sensitive to permissioning changes (especially duplicate databases).
A Crowdstrike style config update that affects all nodes but obviously isn't tested in any QA or staged rollout strategy beforehand (the application panicking straight away with this new file basically proves this).
Finally an error with bot management config files should probably disable bot management vs crash the core proxy.
I'm interested here why they even decided to name Clickhouse as this error could have been caused by any other database. I can see though the replicas updating causing flip / flopping of results would have been really frustrating for incident responders.
Right but also this is a pretty common pattern in distributed systems that publish from databases (really any large central source of truth); it might be like the problem in systems like this. When you're lucky the corner cases are obvious; in the big one we experienced last year, a new row in our database tripped an if-let/mutex deadlock, which our system dutifully (and very quickly) propagated across our entire network.
The solution to that problem wasn't better testing of database permutations or a better staging environment (though in time we did do those things). It was (1) a watchdog system in our proxies to catch arbitrary deadlocks (which caught other stuff later), (2) segmenting our global broadcast domain for changes into regional broadcast domains so prod rollouts are implicitly staged, and (3) a process for operators to quickly restore that system to a known good state in the early stages of an outage.
(Cloudflare's responses will be different than ours, really I'm just sticking up for the idea that the changes you need don't follow obviously from the immediate facts of an outage.)
Cloudflare Access is still experiencing weird issues for us (it’s asking users to SSO login to our public website even though our zone rules - set on a completely different zone - haven’t changed).
I don’t think the infrastructure has been as fully recovered as they think yet…
People really like to hate on Rust for some reason. This wasn’t a Rust problem, no language would have saved them from this kind of issue. In fact, the compiler would have warned that this was a possible issue.
I get it, don’t pick languages just because they are trendy, but if any company’s use case is a perfect fit for Rust it’s cloudflare.
Yeah even if you handled this situation without unwrap() if you just went down an error path that didn't panic, the service would likely still be inoperable if every single request went down the error path.
The reason why people are criticizing is because Rust evangelicals say stuff like "if it compiles it works" or talk about how Rust's type system is so much better than other languages that it catches logic errors like this. You won't see Go or Java developers making such strong claims about their preferred languages.
The real take away is that so much functionality depends on a few players. This is a fundamental flaw in design that is getting worse by the year as the winner takes all winners win. Not saying they didn’t earn their wins. But the fact remains. The system is not robust. Then again, so what. It went down for a while. Maybe we shouldn’t depend on the internet being “up” all the time.
"Customers on our old proxy engine, known as FL, did not see errors, but bot scores were not generated correctly, resulting in all traffic receiving a bot score of zero."
This simply means, the exception handling quality of your new FL2 is non-existent and is not at par / code logic wise similar to FL.
I hope it was not because of AI driven efficiency gains.
In most domains, silently returning 0 in a case where your logic didn't actually calculate the thing you were trying to calculate is far worse than giving a clear error.
A lot of outages off late seem to be related to automated config management.
Companies seem to place a lot of trust is configs being pushed automatically without human review into running systems. Considering how important these configs are, shouldn't they perhaps first be deployed to a staging/isolated network for a monitoring window before pushing to production systems?
Not trying to pontificate here, these systems are more complicated than anything I have maintained. Just trying to think of best practices perhaps everyone can adopt.
Why call .unwrap() in a function which returns Result<_,_>?
For something so critical, why aren't you using lints to identify and ideally deny panic inducing code. This is one of the biggest strengths of using Rust in the first place for this problem domain.
Fly writes a lot of Rust, do you allow `unwrap()` in your production environment? At Modal we only allow `expect("...")` and the message should follow the recommended message style[1].
I'm pretty surprised that Cloudflare let an unwrap into prod that caused their worst outage in 6 years.
After The Great If-Let Outage Of 2024, we audited all our code for that if-let/rwlock problem, changed a bunch of code, and immediately added a watchdog for deadlocks. The audit had ~no payoff; the watchdog very definitely did.
I don't know enough about Cloudflare's situation to confidently recommend anything (and I certainly don't know enough to dunk on them, unlike the many Rust experts of this thread) but if I was in their shoes, I'd be a lot less interested in eradicating `unwrap` everywhere and more in making sure than an errant `unwrap` wouldn't produce stable failure modes.
But like, the `unwrap` thing is all programmers here have to latch on to, and there's a psychological self-soothing instinct we all have to seize onto some root cause with a clear fix (or, better yet for dopaminergia, an opportunity to dunk).
A thing I really feel in threads like this is that I'd instinctively have avoided including the detail about an `unwrap` call --- I'd have worded that part more ambiguously --- knowing (because I have a pathological affinity for this community) that this is exactly how HN would react. Maybe ironically, Prince's writing is a little better for not having dodged that bullet.
Sounds like if nothing else, additional attention around (their?) use of unwrap() is still warranted from where you're sitting then though, no? I don't think there's anything wrong with flagging that.
It's one thing to not want to be the one to armchair it, but that doesn't mean that one has to suppress their normal and obvious reactions. You're allowed to think things even if they're kitsch, you too are human, and what's kitsch depends and changes. Applies to everyone else here by extension too.
Fair. I agree that saying "it's the unwrap" and calling it a day is wrong. Recently actually we've done an exercise on our Worker which is "assume the worst kind of panic happens. make the Worker be ok with it".
But I do feel strongly that the expect pattern is a highly useful control and that naked unwraps almost always indicate a failure to reason about the reliability of a change. An unwrap in their core proxy system indicates a problem in their change management process (review, linting, whatever).
Rust has debug asserts for that. Using expect with a comment about why the condition should not/can't ever happen is idiomatic for cases where you never expect an Err.
This reads to me more like the error type returned by append with names is not (ErrorFlags, i32) and wasn't trivially convertible into that type so someone left an unwrap in place on an "I'll fix it later" basis, but who knows.
Oh absolutely, that's how it would have been treated.
Surely a unwrap_or_default() would have been a much better fit--if fetching features fails, continue processing with an empty set of rules vs stop world.
I am one of those old grey beards (or at least, I got started shipping C code in the 1990s), and I'd leave asserts in prod serverside code given the choice; better that than a totally unpredictable error path.
I don't think "implicitly panicked" is an accurate description since unwrap()'s entire reason for existing is to panic if you unwrap an error condition. If you use unwrap(), you're explicitly opting into the panicking behavior.
I suppose another way to think about it is that Result<T, E> is somewhat analogous to Java's checked exceptions - you can't get the T out unless you say what to do in the case of the E/checked exception. unwrap() in this context is equivalent to wrapping the checked exception in a RuntimeException and throwing that.
Yes it's meant to be used in test code. If you're sure it can't fail do then use .expect() that way it shows you made a choice and it wasn't just a dev oversight.
Limits in systems like these are generally good. They mention the reasoning around it explicitly. It just seems like the handling of that limit is what failed and was missed in review.
The outage sucked for everyone. The root cause also feels like something they could have caught much earlier in a canary rollout from my reading of this.
All that said, to have an outage reported turned around practically the same day, that is this detailed, is quite impressive. Here's to hoping they make their changes from this learning, and we don't see this exact failure mode again.
On 18 November 2025 at 11:20 UTC (all times in this blog are UTC), Cloudflare's network began experiencing significant failures
As of 17:06 all systems at Cloudflare were functioning as normal
May I just say that Matthew Prince is the CEO of Cloudflare and a lawyer by training (and a very nice guy overall). The quality of this postmortem is great but the fact that it is from him makes one respect the company even more.
I don't get why that SQL query was even used in the first place. It seems it fetches feature names at runtime instead of using a static hardcoded schema. Considering this decides the schema of a global config, I don't think the dynamicity is a good idea.
kudos to getting this blog post out so fast, it’s well written and is appreciated.
i’m a little confused on how this was initially confused for an attack though?
is there no internal visibility into where 5xx’s are being thrown? i’m surprised there isn’t some kind of "this request terminated at the <bot checking logic>" error mapping that could have initially pointed you guys towards that over an attack.
also a bit taken aback that .unwrap()’s are ever allowed within such an important context.
Question: customer having issues also couldn't switch their dns to bypass the service, why is the control plane updated along the data plane here it seem a lot of use could save business continuity if they could change their dns entry temporarily
Hold up ,- when I used a C or similar language for accessing a database and wanted to clamp down on memory usage to deterministically control how much I want to allocated, I would explicitly limit the number of rows in the query.
There never was an unbound "select all rows from some table" without a "fetch first N rows only" or "limit N"
If you knew that this design is rigid, why not leverage the query to actually do it ?
Because nothing forced them to and they didn't think of it. Maybe the people writing the code that did the query knew that the tables they were working with never had more than 60 rows and figured "that's small" so they didn't bother with a limit. Maybe the people who wrote the file size limit thought "60 rows isn't that much data" and made a very small file size limit and didn't coordinate with the first people.
Anyway regardless of which language you use to construct a SQL query, you're not obligated to put in a max rows
I imagine there's numerous ways to protect against it and protection should've been added by whoever decided on this optimization. In data layer, create some kind of view which never returns more than 200 rows from base table(s). In code, use some kind of iterator. I'm not a Rust guy, just a C defensive practices type of dude, but maybe they just missed a biggie during a code review.
Cloudflare’s write-up is clear and to the point. A small change spread wider than expected, and they explained where the process failed. It’s a good reminder that reliability depends on strong workflows as much as infrastructure.
> Throwing us off and making us believe this might have been an attack was another apparent symptom we observed: Cloudflare’s status page went down. The status page is hosted completely off Cloudflare’s infrastructure with no dependencies on Cloudflare.
also cloudflare:
> The Cloudflare Dashboard was also impacted due to both Workers KV being used internally and Cloudflare Turnstile being deployed as part of our login flow.
thanks for clarifying! i guess then they never explained why the status page went down, even though it's supposed to be running on independent infrastructure.
Yes, that was missing (along with the London WARP thing). Other comments mentioned that their status page is an Atlassian Statuspage solution, hosted on AWS CloudFront.
Unclear to me if it's an Atlassian-managed deployment they have, or if it's self-managed, I'm not familiar with Statuspage and their website isn't helping. Though if it's managed, I'm not sure how they can know for sure there's no interdependence. (Though I guess we could technically keep that rabbit hole going indefinitely.)
Would be nice if their Turnstile could be turned off on their login page when something like this happens, so we can attempt to route traffic away from Cloudflare during the outage. Or at least have a simple app where this can be modified from.
Honestly... everyone shit themselves that internet doesn't work, but next week this outage will be forgotten by 99% of population. I was doing something on my PC when I saw clear information that Cloudflare is down, so I decided to just go take a nap, then read a book, then go for a walk. Once I was done, the internet was working again. Panic was not necessary on my side.
What I'm trying to say is that things would be much better if everyone took a chill pill and accepted the possibility that in rare instances, the internet doesn't work and that's fine. You don't need to keep scrolling TikTok 24/7.
Wondering why they didn’t disable the bot management temporarily to recover. Websites could have survived temporarily without it compared to the outage itself.
unwraps are so very easy to use and they have bit me so many times because you can nearly never run into a problem and suddenly crashes from an unwrap that almost always was fine
If you deploy a change to your system, and things start to go wrong that same day, the prime suspect (no matter how unlikely it might seem) should be the change you made.
>Currently that limit is set to 200, well above our current use of ~60 features. Again, the limit exists because for performance reasons we preallocate memory for the features.
So they basically hardcoded something, didn't bother to cover the overflow case with unit tests, didn't have basic error catching that would fallback and send logs/alerts to their internal monitoring system and this is why half of the internet went down?
Wow. What a post mortem.
Rather than Monday morning quarterbacking how many ways this could have been prevented, I'd love to hear people sound-off on things that unexpectedly broke. I, for one, did not realize logging in to porkbun to edit DNS settings would become impossible with a cloudflare meltdown
That's unfortunate. I'll need to investigate whether Porkbun plans on decoupling its auth from being reliant on CloudFlare, otherwise I will need to migrate a few domains off of that registrar.
While it's certainly worthwhile to discuss the Technical and Procedural elements that contributed to this Service Outage, the far more important (and mutually-exclusive aspect) to discuss should be:
Why have we built / permitted the building of / Subscribed to such a Failure-intolerant "Network"?
Who's "we"? This is not a trick question, what specific people do you think acted wrongly here? I don't use Cloudflare personally. I don't run any of the sites that do use it. The people who did make the decision to put thier websites behind Cloudflare could stop, and maybe some will, but presumably they're paying for it because they think, perhaps accurately, that they get value out of it. Should some power compel them not to use Cloudflare?
Here's a random post from their blog by the same author from 2017 with an em dash:
> As we wrote before, we believe Blackbird Tech's dangerous new model of patent trolling — where they buy patents and then act their own attorneys in cases — may be a violation of the rules of professional ethics.
tl;dr
A permissions change in a ClickHouse database caused a query to return duplicate rows for a “feature file” used by Cloudflares Bot Management system, which doubled the file size. That oversized file was propagated to their core proxy machines, triggered an unhandled error in the proxy’s bot-module (it exceeded its pre-allocated limit), and as a result the network started returning 5xx errors. The issue wasn’t a cyber-attack — it was a configuration/automation failure.
Excellent write up. Cybersecurity professionals read the story and learn. It’s textbook lesson in post-mortem incident analysis - a mvp for what is expected from us all in a similar situation.
Reputationally this is extremely embarrassing for Cloudflare, but imo they seem to get their feet back on the ground. I was surprised to see not just one, but two apologies to the internet. This just cements how professional and dedicated the Cloudflare team is to ensure stable resilient internet and how embarrassed they must have been.
A reputational hit for sure, but outcome is lessons learned and hopefully stronger resilience.
So an unhandled error condition after an configuration update similar to Crowdstrike - if they had just used a programming language where this can't happen due to the superior type system such as Rust. Oh wait.
> Instead, it was triggered by a change to one of our database systems' permissions which caused the database to output multiple entries into a “feature file” used by our Bot Management system.
And here is the query they used ** (OK, so it's not exactly):
SELECT * from feature JOIN permissions on feature.feature_type_id = permissions.feature_type_id
someone added a new row to permissions and the JOIN started returning two dupe feature rows for each distinct feature.
** "here is the query" is used for dramatic effect. I have no knowledge of what kind of database they are even using much less queries (but i do have an idea).
more edits: OK apparently it's described later in the post as a query against clickhouse's table metadata table, and because users were granted access to an additional database that was actually the backing store to the one they normally worked with, some row level security type of thing doubled up the rows. Not sure why querying system.columns is part of a production level query though, seems overly dynamic.
I can never get used to the error happening at call site rather than within the function where the early return of Err happened. It is not "much cleaner", you have no idea which line and file caused it at call site. By default Returning should have a way of setting a marker which can then be used to map back to the line() and file(). 10+ years and still no ergonomics.
> a change to one of our database systems' permissions which caused the database to output multiple entries into a “feature file” used by our Bot Management system ... to keep [that] system up to date with ever changing threats
> The software had a limit on the size of the feature file that was below its doubled size. That caused the software to fail
A configuration error can cause internet-scale outages. What an era we live in
Edit: also, after finishing my reading, I have to express some surprise that this type of error wasn't caught in a staging environment. If the entire error is that "during migration of ClickHouse nodes, the migration -> query -> configuration file pipeline caused configuration files to become illegally large", it seems intuitive to me that doing this same migration in staging would have identified this exact error, no?
I'm not big on distributed systems by any means, so maybe I'm overly naive, but frankly posting a faulty Rust code snippet that was unwrapping an error value without checking for the error didn't inspire confidence for me!
It would have been caught only in stage if there was similar amount of data in the database. If stage has 2x less data it would have never occurred there. Not super clear how easy it would have been to keep stage database exactly as production database in terms of quantity and similarity of data etc.
I think it's quite rare for any company to have exact similar scale and size of storage in stage as in prod.
But now consider how much extra data Cloudflare at its size would have to have just for staging, doubling or more their costs to have stage exactly as production. They would have to simulate similar amount of requests on top of themselves constantly since presumably they have 100s or 1000s of deployments per day.
In this case it seems the database table in question seemed modest in size (the features for ML) so naively thinking they could have kept stage features always in sync with prod at the very least, but could be they didn't consider that 55 rows vs 60 rows or similar could be a breaking point given a certain specific bug.
It is much easier to test with 20x data if you don't have the amount of data cloudflare probably handles.
That just means it takes longer to test. It may not be possible to do it in a reasonable timeframe with the volumes involved, but if you already have 100k servers running to serve 25M requests per second, maybe briefly booting up another 100k isn’t going to be the end of the world?
Either way, you don’t need to do it on every commit, just often enough that you catch these kinds of issues before they go to prod.
The speed and transparency of Cloudflare publishing this port mortem is excellent.
I also found the "remediation and follow up" section a bit lacking, not mentioning how, in general, regressions in query results caused by DB changes could be caught in future before they get widely rolled out.
Even if a staging env didn't have a production-like volume of data to trigger the same failure mode of a bot management system crash, there's also an opportunity to detect that something has gone awry if there were tests that the queries were returning functionally equivalent results after the proposed permission change. A dummy dataset containing a single http_requests_features column would suffice to trigger the dupe results behaviour.
In theory there's a few general ways this kind of issue could be detected, e.g. someone or something doing a before/after comparison to test that the DB permission change did not regress query results for common DB queries, for changes that are expected to not cause functional changes in behaviour.
Maybe it could have been detected with an automated test suite of the form "spin up a new DB, populate it with some curated toy dataset, then run a suite of important queries we must support and check the results are still equivalent (after normalising row order etc) to known good golden outputs". This style of regression testing is brittle, burdensome to maintain and error prone when you need to make functional changes and update what then "golden" outputs are - but it can give a pretty high probability of detecting that a DB change has caused unplanned functional regressions in query output, and you can find out about this in a dev environment or CI before a proposed DB change goes anywhere near production.
They only recently rewrote their core in Rust (https://blog.cloudflare.com/20-percent-internet-upgrade/) -- given the newness of the system and things like "Over 100 engineers have worked on FL2, and we have over 130 modules" I won't be surprised for further similar incidents.
So they made a newbie mistake in SQL that would not even pass an AI review. They did not verify the change in a test environment. And I guess the logs are so full of errors it is hard to pinpoint which matters. Yikes.
The internet hasn't been the internet in years. It was originally built to withstand wars. The whole idea of our IP based internet was to reroute packages should networks go down. Decentralisation was the mantra and how it differed from early centralised systems such as AOL et al.
This is all gone. The internet is a centralised system in the hand of just a few companies. If AWS goes down half the internet does. If Azure, Google Cloud, Oracle Cloud, Tencent Cloud or Alibaba Cloud goes down a large part of the internet does.
Yesterday with Cloudflare down half the sites I tried gave me nothing but errors.
this is where change management really shines because in a change management environment this would have been prevented by a backout procedure and it would never have been rolled out to production before going into QA, with peer review happening before that... I don't know if they lack change management but it's definitely something to think about
i think that is data rather than code which is where it falls short, in a way you need stringent code and more safeguarded code; it's like if everyone sends you 64k posts as that's all your proxy layer lets in, someone checked sending 128kb and it gave an error before reaching your app - and then someone sends 128kb and the proxy layer has changed - and your app crashes as it was more than 64kb and your app had an assert against that. to actually track issues with erraneous data that overflows well and stuff isn't so much code test but more like fuzz testing, brute force testing etc. which i think people should do; but that's more like we need strong test networks, and also those test networks may need to be more internet like to reflect real issues too, so the whole testing infrastructure in itself becomes difficult to get right - like they have their own tunneling system etc, they could segregate some of their servers and make a test system with better error diagnosis etc potentially. but to my mind, if they had better error propogation back that really identified what was happening and where then that would be a lot better in general. sure, start doing that on a test network. this is something i've beeen tihnking about in general - i made a simple rpc system for being able to send real time rust tracing logs (it allows to just use the normal tracing framework and use a thin rpc layer) back from multiple end servers but that's mostly for granular debugging. i've never quite understood why systems like systemd-journald aren't more network centric when they're going to be big and complex kitchensink approaches - apparently there's dbus support, but to my mind something inbetween debugging level of code and warning/info. like even if it's doing things like 1/20 of log info it's too much volume if things like large files getting close to limits is increasing etc and we can see this as things run, and can see if it's localised or common etc it'd help have more resilient systems. something may already exist in this line but i didn't come across anything in a reasonably passive way - i mean there's debugging tools like dtrace etc that have been around for ages.
28M 500 errors/sec for several hours from a single provider. Must be a new record.
No other time in history has one single company been responsible for so much commerce and traffic. I wonder what some outage analogs to the pre-internet ages would be.
Something like a major telco going out, for example the AT&T 1990 outage of long distance calling:
> The standard procedures the managers tried first failed to bring the network back up to speed and for nine hours, while engineers raced to stabilize the network, almost 50% of the calls placed through AT&T failed to go through.
> Until 11:30pm, when network loads were low enough to allow the system to stabilize, AT&T alone lost more than $60 million in unconnected calls.
> Still unknown is the amount of business lost by airline reservations systems, hotels, rental car agencies and other businesses that relied on the telephone network.
Yes, all(most) eggs should not be in one basket. Perfect opportunity to setup a service that checks cloudflare then switches a site's DNS to akami as a backup.
Absolute volume maybe[1], as relative % of global digital communication traffic, the era of early telegraph probably has it beat.
In the pre digital era, East India Company dwarfs every other company in any metric like commerce controlled, global shipping, communication traffic, private army size, %GDP , % of workforce employed by considerable margins.
The default was large consolidated organization throughout history, like say Bell Labs, or Standard Oil before that and so on, only for a brief periods we have enjoyed benefits of true capitalism.
[1] Although I suspect either AWS or MS/Azure recent down-times in the last couple of years are likely higher
Best post mortem I've read in a while, this thing will be studied for years.
A bit ironic that their internal FL2 tool is supposed to make Cloudflare "faster and more secure" but brought a lot of things down. And yeah, as other have already pointed out, that's a very unsafe use of Rust, should've never made it to production.
This is the first significant outage that has involved Rust code, and as you can see the .unwrap is known to carry the risk of a panic and should never be used on production code.
I think you should give me a credit for all the income I lost due to this outage. Who authorized a change to the core infrastructure during the period of the year when your customers make the most income? Seriously, this is a management failure at the highest levels of decision-making. We don't make any changes to our server infrastructure/stack during the busiest time of the year, and neither should you. If there were an alternative to Cloudflare, I'd leave your service and move my systems elsewhere.
I think you should get exactly what the contract you signed said you'd get. Outages happen in all infrasturture. Planned and unplanned ones both. The SLA and SLO are literal acknowledgements of the fact that and part of the contract for that reason.
Its fair to be upset at their decision making - use that to renegotiate your contract.
Did some $300k chief of IT blame it all on some overworked secretary clicking a link in an email they should have run through a filter? Because that’s the MO.
> The change explained above resulted in all users accessing accurate metadata about tables they have access to. Unfortunately, there were assumptions made in the past, that the list of columns returned by a query like this would only include the “default” database:
SELECT
name,
type
FROM system.columns
WHERE
table = 'http_requests_features'
order by name;
Note how the query does not filter for the database name. With us gradually rolling out the explicit grants to users of a given ClickHouse cluster, after the change at 11:05 the query above started returning “duplicates” of columns because those were for underlying tables stored in the r0 database.
Here is a bit more context in addition to the quote above. A ClickHouse permissions change made a metadata query start returning duplicate column metadata from an extra schema, which more than doubled the size and feature count of a Bot Management configuration file. When this oversized feature file was deployed to edge proxies, it exceeded a 200-feature limit in the bot module, causing that module to panic and the core proxy to return 5xx errors globally
- Their database permissions changed unexpectedly (??)
- This caused a 'feature file' to be changed in an unusual way (?!)
- Their SQL query made assumptions about the database; their permissions change thus resulted in queries getting additional results, permitted by the query
- Changes were propagated to production servers which then crashed those servers (meaning they weren't tested correctly)
- They hit an internal application memory limit and that just... crashed the app
- The crashing did not result in an automatic backout of the change, meaning their deployments aren't blue/green or progressive
- After fixing it, they were vulnerable to a thundering herd problem
- Customers who were not using bot rules were not affected; CloudFlare's bot-scorer generated a constant bot score of 0, meaning all traffic is bots
In terms of preventing this from a software engineering perspective, they made assumptions about how their database queries work (and didn't validate the results), and they ignored their own application limits and didn't program in either a test for whether an input would hit a limit, or some kind of alarm to notify the engineers of the source of the problem.
From an operations perspective, it would appear they didn't test this on a non-production system mimicing production; they then didn't have a progressive deployment; and they didn't have a circuit breaker to stop the deployment or roll-back when a newly deployed app started crashing.
People jump to say things like "where's the rollback" and, like, probably yeah, but keep in mind that speculative rollback features (that is: rollbacks built before you've experienced the real error modes of the system) are themselves sources of sometimes-metastable distributed system failures. None of this is easy.
How about where's the most basic test to check if your config file will actually run at all in your application? It was a hard-coded memory limit; a git-hook test suite run a MacBook would have caught this. But nooo, let's not run the app for 0.01 seconds with this config before sending it out to determine the fate of the internet?
This is literally the CrowdStrike bug, in a CDN. This is the most basic, elementary, day 0 test you could possibly invent. Forget the other things they fucked up. Their app just crashes with a config file, and nobody evaluates it?! Not every bug is preventable, but an egregious lack of testing is preventable.
This is what a software building code (like the electrical code's UL listings that prevent your house from burning down from untested electrical components) is intended to prevent. No critical infrastructure should be legal without testing, period.
just before this outage i was exploring bunnycdn as the idea of cloudflare taking over dns still irks me slightly. there are competitors. but there's a certain amount of scale that cloudflare offers which i think can help performance in general. that said in the past i found cloudflare performance terrible when i was doings lots of testing. they are predominantly a pull based system not a push, so if content isn't current the cache miss performance can be kind of blah. i think their general backhaul paths have improved, but at least from new zealand they used to seem to do worse than hitting a los angeles proxy that then hits origin. (although google was in a similar position before, where both 8.8.8.8 and www.google.co.nz/.com were both faster via los angeles than via normal paths - i think google were doing asia parent, like if testing 8.8.8.8 misses it was super far away). i think now that we have http/3 etc though that performance is a bit simpler to achieve, and that ddos, bot protection is kind of the differentiator, and i think that cloudflare's bot protection may work reasonably well in general?
This is the multi-million dollar .unwrap() story. In a critical path of infrastructure serving a significant chunk of the internet, calling .unwrap() on a Result means you're saying "this can never fail, and if it does, crash the thread immediately."The Rust compiler forced them to acknowledge this could fail (that's what Result is for), but they explicitly chose to panic instead of handle it gracefully. This is textbook "parse, don't validate" anti-pattern.
I know, this is "Monday morning quarterbacking", but that's what you get for an outage this big that had me tied up for half a day.
I’ve led multiple incident responses at a FAANG, here’s my take. The fundamental problem here is not Rust or the coding error. The problem is:
1. Their bot management system is designed to push a configuration out to their entire network rapidly. This is necessary so they can rapidly respond to attacks, but it creates risk as compared to systems that roll out changes gradually.
2. Despite the elevated risk of system wide rapid config propagation, it took them 2 hours to identify the config as the proximate cause, and another hour to roll it back.
SOP for stuff breaking is you roll back to a known good state. If you roll out gradually and your canaries break, you have a clear signal to roll back. Here was a special case where they needed their system to rapidly propagate changes everywhere, which is a huge risk, but didn’t quite have the visibility and rapid rollback capability in place to match that risk.
While it’s certainly useful to examine the root cause in the code, you’re never going to have defect free code. Reliability isn’t just about avoiding bugs. It’s about understanding how to give yourself clear visibility into the relationship between changes and behavior and the rollback capability to quickly revert to a known good state.
Cloudflare has done an amazing job with availability for many years and their Rust code now powers 20% of internet traffic. Truly a great team.
The "coding error" is a somewhat deliberate choice to fail eagerly that is usually safe but doesn't align with the need to do something (propagation of the configuration file) without failing.
I'm sure that there are misapplied guidelines to do that instead of being nice to incoming bot management configuration files, and someone might have been scolded (or worse) for proposing or attempting to handle them more safely.
Does their ring based rollout really truly have to be 0->100% in a few seconds?
I don’t really buy this requirement. At least make it configurable with a more reasonable default for “routine” changes. E.g. ramping to 100% over 1 hour.
As long as that ramp rate is configurable, you can retain the ability to respond fast to attacks by setting the ramp time to a few seconds if you truly think it’s needed in that moment.
I think defence against a DDOS against your network is the best reason for a quick rollout
Maybe, but in that case maybe have some special casing logic to detect that yes indeed we're under a massive DDOS at this very moment, do a rapid rollout of this thing that will mitigate said DDOS. Otherwise use the default slower one?
Of course, this is all so easy to say after the fact..
Isn’t CF under a ‘massive DDOS’ 24/7 pretty much by definition? When does malicious traffic rest, and how many targets of same aren’t using CF?
Exactly the right take. Even when you want to have rapid changes on your infra, do it at least by region. You can start with the region where the least amount of users are impacted and if everything is fine, there is no elevated number of crashes for example, you can move forward. It was a standard practice at $RANDOM_FAANG when we had such deployments.
Rolling out new code should be done differently than rolling out new data to fight bots.
If every time there's a new bot someone needs to write code that can blow up their whole service, maybe they need to iterate a bit on this design?
Thanks for this assessment.
In a productive way, this view also shifts the focus to improving the system (visibility etc), empowering the team, rather than focusing on the code which broke (probably strikes fear in the individuals, to do anything!)
Partial disagree. There should be lints against 'unwrap's. An 'expect' at least forces you to write down why you are so certain it can't fail. An unwrap is not just hubris, it's also laziness, and has no place in sensitive code.
And yes, there is a lint you can use against slicing ('indexing_slicing') and it's absolutely wild that it's not on by default in clippy.
This guy SREs
> Their bot management system is designed to push a configuration out to their entire network rapidly.
Once every 5m is not "rapidly". It isn't uncommon for configuration systems to do it every few seconds [0].
> While it’s certainly useful to examine the root cause in the code.
Believe the issue is as much an output from a periodic run (clickhouse query) caused by (on the surface, an unrelated change) causing this failure. That is, the system that validated the configuration (FL2) was different to the one that generated it (ML Bot Management DB).
Ideally, it is the system that vends a complex configuration that also vends & tests the library to consume it, or the system that consumes it, does so as if it was "tasting" the configuration first before devouring it unconditionally [1].
Of course, as with all distributed system failures, this is all easier said and done in hindsight.
[0] Avoiding overload in distributed systems by putting the smaller service in control (pg 4), https://d1.awsstatic.com/builderslibrary/pdfs/Avoiding%20ove...
[1] Lessons from CloudFront (2016), https://youtube.com/watch?v=n8qQGLJeUYA&t=1050
By rapid I mean a rapid rollout of changes to 100% of the fleet, not how often changes are made.
>Once every 5m is not "rapidly".
Isn't rapidly more of how long it takes to get from A to Z rather than how often it is performed? You can push out a configuration update every fortnight but if it goes through all of your global servers in three seconds, I'd call it quite rapid.
It seems people have a blind spot for unwrap, perhaps because it's so often used in example code. In production code an unwrap or expect should be reviewed exactly like a panic.
It's not necessarily invalid to use unwrap in production code if you would just call panic anyway. But just like every unsafe block needs a SAFETY comment, every unwrap in production code needs an INFALLIBILITY comment. clippy::unwrap_used can enforce this.
Yes, I always thought it was wrong to use unwrap in examples. I know, people want to keep examples simple, but it trains developers to use unwrap() as they see that everywhere. Yes, there are places where it's ok as that blog post explains so well: https://burntsushi.net/unwrap/ But most devs IMHO don't have the time to make the call correctly most of the time... so it's just better to do something better, like handle the error and try to recover, or if impossible, at least do `expect("damn it, how did this happen")`.
Dunno, I think the alternatives have their own pretty significant downsides. All would require front loading more in-depth understanding of error handling and some would just be quite a bit more verbose.
IMO making unwrap a clippy lint (or perhaps a warning) would be a decent start. Or maybe renaming unwrap.
This thread warms my heart. Rust has set a new baseline that many and myself now take for granted.
We are now discussing what can be done to improve code correctness beyond memory and thread safety. I am excited for what is to come.
Alternatively you can look at actually innovative programming languages to peek at the next 20 years of innovation.
I am not sure that watching the trendy forefront successfully reach the 1990s and discuss how unwrapping Option is potentially dangerous really warm my heart. I can’t wait for the complete meltdown when they discover effect systems in 2040.
To be more serious, this kind of incident is yet another reminder that software development remains miles away from proper engineering and even key providers like Cloudfare utterly fail at proper risk management.
Celebrating because there is now one popular language using static analysis for memory safety feels to me like being happy we now teach people to swim before a transatlantic boat crossing while we refuse to actually install life boats.
To me the situation has barely changed. The industry has been refusing to put in place strong reliability practices for decades, keeps significantly under investing in tools mitigating errors outside of a few fields where safety was already taken seriously before software was a thing and keeps hiding behind the excuse that we need to move fast and safety is too complex and costly while regulation remains extremely lenient.
I mean this Cloudfare outage probably cost millions of dollars of damage in aggregate between lost revenue and lost productivity. How much of that will they actually have to pay?
Let's try to make effect systems happen quicker than that.
> I mean this Cloudfare outage probably cost millions of dollars of damage in aggregate between lost revenue and lost productivity. How much of that will they actually have to pay?
Probably nothing, because most paying customers of cloudflare are probably signing away their rights to sue Cloudflare for damages by being down for a while when they purchase Cloudflare's services (maybe some customers have SLAs with monetary values attached, I dunno). I honestly have a hard time suggesting that those customers are individually wrong to do so - Cloudflare isn't down that often, and whatever amount it cost any individual customer by being down today might be more than offset by the DDOS protection they're buying.
Anyway if you want Cloudflare regulated to prevent this, name the specific regulations you want to see. Should it be illegal under US law to use `unwrap` in Rust code? Should it be illegal for any single internet services company to have more than X number of customers? A lot of the internet also breaks when AWS goes down because many people like to use AWS, so maybe they should be included in this regulatory framework too.
> every unwrap in production code needs an INFALLIBILITY comment. clippy::unwrap_used can enforce this.
How about indexing into a slice/map/vec? Should every `foo[i]` have an infallibility comment? Because they're essentially `get(i).unwrap()`.
Yes? Funnily enough, I don't often use indexed access in Rust. Either I'm looping over elements of a data structure (in which case I use iterators), or I'm using an untrusted index value (in which case I explicitly handle the error case). In the rare case where I'm using an index value that I can guarantee is never invalid (e.g. graph traversal where the indices are never exposed outside the scope of the traversal), then I create a safe wrapper around the unsafe access and document the invariant.
If that's the case then hats off. What you're describing is definitely not what I've seen in practice. In fact, I don't think I've ever seen a crate or production codebase that documents infallibility of every single slice access. Even security-critical cryptography crates that passed audits don't do that. Personally, I found it quite hard to avoid indexing for graph-heavy code, so I'm always on the lookout for interesting ways to enforce access safety. If you have some code to share that would be very interesting.
> I don't think I've ever seen a crate or production codebase that documents infallibility of every single slice access.
The smoltcp crate typically uses runtime checks to ensure slice accesses made by the library do not cause a panic. It's not exactly equivalent to GP's assertion, since it doesn't cover "every single slice access", but it at least covers slice accesses triggered by the library's public API. (i.e. none of the public API functions should cause a panic, assuming that the runtime validation after the most recent mutation succeeds).
Example: https://docs.rs/smoltcp/latest/src/smoltcp/wire/ipv4.rs.html...
My rule of thumb is that unchecked access is okay in scenarios where both the array/map and the indices/keys are private implementation details of a function or struct, since an invariant is easy to manually verify when it is tightly scoped as such. I've seen it used it in:
* Graph/tree traversal functions that take a visitor function as a parameter
* Binary search on sorted arrays
* Binary heap operations
* Probing buckets in open-addressed hash tables
> graph-heavy code
Could you share some more details, maybe one fully concrete scenario? There are lots of techniques, but there's no one-size-fits-all solution.
Sure, these days I'm mostly working on a few compilers. Let's say I want to make a fixed-size SSA IR. Each instruction has an opcode and two operands (which are essentially pointers to other instructions). The IR is populated in one phase, and then lowered in the next. During lowering I run a few peephole and code motion optimizations on the IR, and then do regalloc + asm codegen. During that pass the IR is mutated and indices are invalidated/updated. The important thing is that this phase is extremely performance-critical.
And it's fine for a compiler to panic when it violates an assumption. Not so with the Cloudflare code under discussion.
Usually you'd want to write almost all your slice or other container iterations with iterators, in a functional style.
For the 5% of cases that are too complex for standard iterators? I never bother justifying why my indexes are correct, but I don't see why not.
You very rarely need SAFETY comments in Rust because almost all the code you write is safe in the first place. The language also gives you the tool to avoid manual iteration (not just for safety, but because it lets the compiler eliminate bounds checks), so it would actually be quite viable to write these comments, since you only need them when you're doing something unusual.
I didn't restate the context from the code we're discussing: it must not panic. If you don't care if the code panics, then go ahead and unwrap/expect/index, because that conforms to your chosen error handling scheme. This is fine for lots of things like CLI tools or isolated subprocesses, and makes review a lot easier.
So: first, identify code that cannot be allowed to panic. Within that code, yes, in the rare case that you use [i], you need to at least try to justify why you think it'll be in bounds. But it would be better not to.
There are a couple of attempts at getting the compiler to prove that code can't panic (e.g., the no-panic crate).
What about memory allocation - how will you stop that from panicking ? `Vec::resize` will always panic in Rust. And this is just one example out of thousands in the Rust stdlib.
Unless the language addresses no-panic in its governing design or allows try-catch, not sure how you go about this.
That is slowly being addressed, but meanwhile it’s likely you have a reliable upper bound on how much heap your service needs, so it’s a much smaller worry. There are also techniques like up-front or static allocation if you want to make more certain.
Yep and this postmortem details how their proxy modules use static allocation.
In TFA they mentioned they preallocate all the memory up front
Indexing is comparatively rare given the existence of iterators, IMO. If your goal is to avoid any potential for panicking, I think you'd have a harder time with arithmetic overflow.
Cargo needs to grow a label for crates that provably do not panic. (Neverminding allocations and things outside our control flow.)
I want to ban crates that panic from my dependency chain.
The language could really use an extra set of static guarantees around this. I would opt in.
> I want to ban crates that panic from my dependency chain.
Which means banning anything that allocates memory and thousands of stdlib functions/methods.
See the immediately preceding sentence.
I'm fine with allocation failures. I don't want stupid unwrap()s, improper slice access, or other stupid and totally preventable behavior.
There are things inside the engineer's control. I want that to not panic.
Your pair of posts is very interesting to me. Can you share with me: What is your programming environment such that you are "fine with allocation failures"? I'm not doubting you, but for me, if I am doing systems programming with C or C++, my program is doomed if a malloc fails! When I saw your post, I immediately thought: Am I doing it wrong? If I get a NULL back from malloc(), I just terminate with an error message.
Not GP but I read "I'm fine with allocation failures" as "I'm OK with my program terminating if it can't allocate (but not for other errors)".
I mean, yeah, if I am using a library, as an user of this library, I would like to be able to handle the error myself. Having the library decide to panic, for example, is the opposite of it.
I think I'd prefer a compile-time guarantee.
Something that allows me to tag annotate a function (or my whole crate) as "no panic", and get a compile error if the function or anything it calls has a reachable panic.
This will allow it to work with many unmodified crates, as long as constant propagation can prove that any panics are unreachable. This approach will also allow crates to provide panicking and non panicking versions of their API (which many already do).
I think the most common solution at the moment is dtolnay's no_panic [0]. That has a bunch of caveats, though, and the ergonomics leave something to be desired, so a first-party solution would probably be preferable.
[0]: https://github.com/dtolnay/no-panic
Yes, I want that. I also want to be able to (1) statically apply a badge on every crate that makes and meets these guarantees (including transitively with that crate's own dependencies) so I can search crates.io for stronger guarantees and (2) annotate my Cargo.toml to not import crates that violate this, so time isn't wasted compiling - we know it'll fail in advance.
On the subject of this, I want more ability to filter out crates in our Cargo.toml. Such as a max dependency depth. Or a frozen set of dependencies that is guaranteed not to change so audits are easier. (Obviously we could vendor the code in and be in charge of our own destiny, but this feels like something we can let crate authors police.)
For iteration, yes. But there's other cases, like any time you have to deal with lots of linked data structures. If you need high performance, chances are that you'll have to use an index+arena strategy. They're also common in mathematical codebases.
I mean... yeah, in general. That's what iterators are for.
It's the same blind spot people have to Java's checked exceptions. People commonly resort to Pokemon exception handling and either blindly ignoring or rethrowing as a runtime exception. When Rust got popular, I was a bit confused by people talking about how great Result it's essentially a checked exception without a stack trace.
"Checked Exceptions Are Actually Good" gang, rise up! :p
I think adoption would have played out very different if there had only been some more syntactic-sugar. For example, an easy syntax for saying: "In this method, any (checked) DeepException e that bubbles up should immediately be replaced by a new (checked) MylayerException(e) that contains the original one as a cause.
We might still get lazy programmers making systems where every damn thing goes into a generic MylayerException, but that mess would still be way easier to fix later than a hundred scattered RuntimeExceptions.
There is also the problem that they decided to make all references nullable, so `NullPointerException`s could appear everywhere. This "forced" them to introduce the escape hatch of `RuntimeException`, which of course was way overused immediately, normalizing it.
I'm with you! Checked exceptions are actually good and the hate for them is super short sighted. The exact same criticisms levied at checked exceptions apply to static typing in general, but people acknowledge the great value static types have for preventing errors at compile time. Checked exceptions have that same value, but are dunked on for some reason.
The dislike is probably because of 2 reasons.
1. in most cases they don't want to handle `InterruptedException` or `IOException` and yet need to bubble them up. In that case the code is very verbose.
2. it makes lambdas and functions incompatible. So eg: if you're passing a function to forEach, you're forced to wrap it in runtime exception.
3. Due to (1) and (2), most people become lazy and do `throws Exception` which negates most advantages of having exceptions in the first place.
In line-of-business apps (where Java is used the most), an uncaught exception is not a big deal. It will bubble up and gets handled somewhere far up the stack (eg: the server logger) without disrupting other parts of the application. This reduces the utility of having every function throw InterruptedException / IOException when those hardly ever happen.
> 2. it makes lambdas and functions incompatible.
This is true, but the hate predated lambdas in Java.
Yeah, in both cases it's a layering situation, where it's the duty of your code to decide what layers of abstraction need to be be bridged, and to execute on that decision. Translating/wrapping exception-types from deeper functions is the same as translating/wrapping return-types the same places.
I think it comes down to a psychological or use-case issue: People hate thinking about errors and handling them, because it's that hard stuff that always consumes more time than we'd like to think. Not just digitally, but in physical machines too. It's also easier to put off "for later."
Checked exceptions in theory were good, but Java simply did not add facilities to handle or support them well in many APIs. Even the new API's in Java - Streams, etc do not support checked exceptions.
> it's essentially a checked exception without a stack trace
In theory, theory and practice are the same. In practice...
You can't throw a checked exception in a stream, this fact actually underlines the key difference between an exception and a Result: Result is in return position and exceptions are a sort of side effect that has its own control flow. Because of that, once your method throws an Exception or you are writing code in a try block that catches an exception, you become blind to further exceptions of that type, even if you might be able to or required to fix those errors. Results are required to be handled individually and you get syntactic sugar to easily back propagate.
It is trivial to include a stack trace, but stack traces are really only useful for identifying where something occurred, and generally what is superior is attaching context as you back propagate which trivially occurs with judicious use of custom error types with From impls. Doing this means that the error message uniquely defines the origin and paths it passed through without intermediate unimportant stack noise. With exceptions you would always need to catch each exception and rethrow a new exception containing the old to add contextual information, then to avoid catching to much you need variables that will be initialized inside the try block defined outside of the try block. So stack traces are basically only useful when you are doing Pokemon exception handling.
It's a lot lighter: a stack trace takes a lot of overhead to generate; a result has no overhead for a failure. The overhead (panic) only comes once the failure can't be handled. (Most books on Java/C# don't explain that throwing exceptions has high performance overhead.)
Exceptions force a panic on all errors, which is why they're supposed to be used in "exceptional" situations. To avoid exceptions when an error is expected, (eof, broken socket, file not found,) you either have to use an unnatural return type or accept the performance penalty of the panic that happens when you "throw."
In Rust, the stack trace happens at panic (unwrap), which is when the error isn't handled. IE, it's not when the file isn't found, it's when the error isn't handled.
> Exceptions force a panic on all errors
What do you mean?
Exceptions do not force panic at all. In most practical situations, an exception unhandled close to where it was thrown will eventually get logged. It's kind of a "local" panic, if you will, that will terminate the specific function, but the rest of the program will remain unaffected. For example, a web server might throw an exception while processing a specific HTTP request, but other HTTP requests are unaffected.
Throwing an exception does not necessarily mean that your program is suddenly in an unsupported state, and therefore does not require terminating the entire program.
> a stack trace takes a lot of overhead to generate
Can't Hotspot not generate the stack trace when it knows the exception will be caught and the stack trace ignored?
checked exceptions failed because when used properly they fossilize method signatures. they're fine if your code will never be changed and they're fine when you control 100% of users of the throwing code. if you're distributing a library... no bueno.
> When Rust got popular, I was a bit confused by people talking about how great Result it's essentially a checked exception without a stack trace.
It's not a checked exception without a stack trace.
Rust doesn't have Java's checked or unchecked exception semantics at the moment. Panics are more like Java's Errors (e.g. OOM error). Results are just error codes on steroids.
Pet peeve: unwrap() should be deprecated and renamed or_panic(). More consistent with the rest of stdlib methods and appropriately scarier.
That's kind of what I'm saying with the blind spot comment. The words "unwrap" and "expect" should be just as much a scary red flag as the word "panic", but for some reason it seems a lot of people don't see them that way.
Even in lowly Java, they later added to Optional the orElseThrow() method since the name of the get() method did not connote the impact of unwrapping an empty Optional.
A lot of stuff should be done about the awful unwrap family of methods.
A few ideas:
- It should not compile in production Rust code
- It should only be usable within unsafe blocks
- It should require explicit "safe" annotation from the engineer. Though this is subject to drift and become erroneous.
- It should be possible to ban the use of unsafe in dependencies and transitive dependencies within Cargo.
Than they are going to write None | Err => yolo() that has the same impact. It is not the syntax or the semantic meaning is the problem here but the fact that there is no monitoring around the elevated error counts after a deployment.
Software engineers tend to get stuck in software problems and thinking that everything should be fixed in code. In reality there are many things outside of the code that you can do to operate unreliable components safely.
The `unsafe` keyword means something specific in Rust, and panicking isn't unsafe by Rust's definition. Sometimes avoiding partial functions just isn't feasible, and an unwrap (or whatever you want to call the method) is a way of providing a (runtime-checked) proof to the compiler that the function is actually total.
Panics should be explicit, not implicit.
unwrap() should effectively work as a Result<> where the user must manually invoke a panic in the failure branch. Make special syntax if a match and panic is too much boilerplate.
This is like an implicit null pointer exception that cannot be statically guarded against.
I want a way to statically block any crates doing this from my dependency chain.
unwrap is explicit.
Not explicit enough, apparently.
Not sure what you're saying with the "work as a Result<>" part...unwrap is a method on Result. I think you're just saying the unwrap/expect methods should be eliminated?
> In production code an unwrap or expect should be reviewed exactly like a panic.
An unwrap should never make it to production IMHO. It's fine while prototyping, but once the project gets closer to production it's necessary to just grep `uncheck` in your code and replace those that can happen with a proper error management and replace those that cannot happen with `expect`, with a clear justification of why they cannot happen unless there's a bug somewhere else.
I would say, sure, if you feel the same way about panic calls making to production. In other words, review all of them the same way. Because writing unwrap/expect is exactly the same as writing “if error, panic”.
> It seems people have a blind spot for unwrap
Not unlike people having a blind spot for Rust in general, no?
Isn't the point of this article that pieces of infrastructure don't go down to root causes, but due to bad combinations of components that are correct individually? After reading "engineering a safer world", I find root cause analysis rather reductionistic, because it wasn't just an unwrap, it was that the payload was larger than normal, because of a query that didn't select by database, because a clickhouse made more databases visible. Hard to say "it was just due to an unwrap" imo. Especially in terms of how to fix an issue going forwards. I think the article lists a lot of good ideas, that aren't just "don't unwrap", like enabling more global kill switches for features, or eliminating the ability for core dumps or other error reports to overwhelm system resources.
You're right. A good postmortem/root cause analysis would START from "unwrap" and continue from there.
You might start with a basic timeline of what happened, then you'd start exploring: why did this change affect so many customers (this would be a line of questioning to find a potential root cause), why did it take so long to discover or recover (this might be multiple lines of questioning), etc.
> This is the multi-million dollar .unwrap() story.
That's too semantic IMHO. The failure mode was "enforced invariant stopped being true". If they'd written explicit code to fail the request when that happened, the end result would have been exactly the same.
>If they'd written explicit code to fail the request when that happened, the end result would have been exactly the same.
If the `.unwrap()` was replaced with `.expect("Feature config is too large!")` it would certainly make the outage shorter.
> If the `.unwrap()` was replaced with `.expect("Feature config is too large!")` it would certainly make the outage shorter.
It wouldn't, not meaningfully. The outage was caused by change in how they processed the queries. They had no way to observe the changes, nor canaries to see that change is killing them. Plus, they would still need to manually feed and restart services that ingested bad configs.
`expect` would shave a few minutes; you would still spend hours figuring out and fixing it.
Granted, using expect is better, but it's not a silver bullet.
In general for unexpected errors like these the internal function should log the error, and I assume it was either logged or they can quickly deduce reason based on the line number.
>> This is the multi-million dollar .unwrap() story.
> That's too semantic IMHO. The failure mode was "enforced invariant stopped being true". If they'd written explicit code to fail the request when that happened, the end result would have been exactly the same.
Problem is, the enclosing function (`fetch_features`) returns a `Result`, so the `unwrap` on line #82 only serves as a shortcut a developer took due to assuming `features.append_with_names` would never fail. Instead, the routine likely should have worked within `Result`.
> Instead, the routine likely should have worked within `Result`.
But it's a fatal error. It doesn't matter whether it's implicit or explicit, the result is the same.
Maybe you're saying "it's better to be explicit", as a broad generalization I don't disagree with that.
But that has nothing to do with the actual bug here, which was that the invariant failed. How they choose to implement checking and failing the invariant in the semantics of the chosen language is irrelevant.
A failed config load probably shouldn't be a fatal error if a valid config is already loaded?
Hard to say. Why would you load a new config if a valid config is already loaded?
Maybe the new config has a new update. Who knows? Do we want to keep operating on the old config? Maybe maybe not.
But operating on old config when you don't want to is definitely worse.
And dare I say, an exhibition of hindsight bias.
semantic? or pedantic?
> Today, many friends pinged me saying Cloudflare was down. As a core developer of the first generation of Cloudflare FL, I'd like to share some thoughts.
> This wasn't an attack, but a classic chain reaction triggered by “hidden assumptions + configuration chains” — permission changes exposed underlying tables, doubling the number of lines in the generated feature file. This exceeded FL2's memory preset, ultimately pushing the core proxy into panic.
> Rust mitigates certain errors, but the complexity in boundary layers, data flows, and configuration pipelines remains beyond the language's scope. The real challenge lies in designing robust system contracts, isolation layers, and fail-safe mechanisms.
> Hats off to Cloudflare's engineers—those on the front lines putting out fires bear the brunt of such incidents.
> Technical details: Even handling the unwrap correctly, an OOM would still occur. The primary issue was the lack of contract validation in feature ingest. The configuration system requires “bad → reject, keep last-known-good” logic.
> Why did it persist so long? The global kill switch was inadequate, preventing rapid circuit-breaking. Early suspicion of an attack also caused delays.
> Why not roll back software versions or restart?
> Rollback isn't feasible because this isn't a code issue—it's a continuously propagating bad configuration. Without version control or a kill switch, restarting would only cause all nodes to load the bad config faster and accelerate crashes.
> Why not roll back the configuration?
> Configuration lacks versioning and functions more like a continuously updated feed. As long as the ClickHouse pipeline remains active, manually rolling back would result in new corrupted files being regenerated within minutes, overwriting any fixes.
https://x.com/guanlandai/status/1990967570011468071
This tweet thread invokes genuine despair in me. Do we really have to outsource even our tweets to LLMs? Really? I mean, I get spambots and the like tweeting mass-produced slop. But what compels a former engineer of the company in question to offer LLM-generated "insight" to the outage? Why? For what purpose?
* For clarity, I am aware that the original tweets are written in Chinese, and they still have the stench of LLM writing all over them; it's not just the translation provided in the above comment.
Out of interest... apart from the em dash, how else can you tell it's an LLM response? What are the telltales signs?
It’s kinda funny that the “not …, but…” + em dash slop signifiers translate directly to Chinese “不是。。。而是。。。” and double-width em dash
To be fair, this failed in the non-rust path too because the bot management returned that all traffic was a bot. But yes, FL2 needs to catch panics from individual components but I’m not sure if failing open is necessarily that much better (it was in this case but the next incident could easily be the result of failing open).
But more generally you could catch the panic at the FL2 layer to make that decision intentional - missing logic at that layer IMHO.
Catching panic probably isn’t a great idea if there’s any unsafe code in the system. (Do the unsafe blocks really maintain heap invariants if across panics?)
Unsafe blocks have nothing to do with it. Yes - they maintain all the same invariants as safe blocks or those unsafe blocks are unsound regardless of panics. But there’s millions of way to architect this (eg a supervisor process that notices which layer in FL2 is crashing and just completely disables that layer when it starts up the proxy again. There’s challenges here because then you have to figure out what constitutes a perma crashing (eg what if it’s just 20% of all sites? Do you disable?). And in the general case you have the fail open/fail close decision anyway which you should just annotate individual layers with.
But the bigger change is to make sure that config changes roll out gradually instead of all at once. That’s the source of 99% of all widespread outages
Incremental config changes sounds like it could lead to a LOT of bugs
Incremental in terms of 1% of the fleet using it, then 5% etc. this is standard course.
Another option is to make sure that config changes that fail to parse continue using the old config instead of resulting in an unusable service.
I think the parent is implying that the panic should be "caught" via a supervisor process, Erlang-style, rather than implying the literal use of `catch_unwind` to resume within the same process.
Supervisor is the brutalist way. But catch_unwind may be needed for perf and other reasons.
But ultimately it’s not the panic that’s the problem but a failure to specify how panics within FL2 layers should be handled; each layer is at least one team and FL2’s job is providing a safe playground for everyone to safely coexist regardless of the misbehavior of any single component
But as always such failures are emblematic of multiple things going wrong at once. You probably want to end up using both catch_unwind for the typical case and the supervisor for the case where there’s a segfault in some unsafe code you call or native library you invoke.
I also mention the fundamental tension of do you want to fail open or closed. Most layers should probably fail open. Some layers (eg auth) it’s safer to fail closed.
This is assuming that the process could have done anything sensible while it had the malformed feature file. It might be in this case that this was one configuration file of several and maybe the program could have been built to run with some defaults when it finds this specific configuration invalid, but in the general case, if a program expects a configuration file and can't do anything without it, panicking is a normal thing to do. There's no graceful handling (beyond a nice error message) a program like Nginx could do on a syntax error in its config.
The real issue is further up the chain where the malformed feature file got created and deployed without better checks.
> panicking is a normal thing to do
I do not think that if the bot detection model inside your big web proxy has a configuration error it should panic and kill the entire proxy and take 20% of the internet with it. This is a system that should fail gracefully and it didn't.
> The real issue
Are there single "real issues" with systems this large? There are issues being created constantly (say, unwraps where there shouldn't be, assumptions about the consumers of the database schema) that only become apparent when they line up.
I don't know too much about how the feature file distribution works but in the event of failure to read a new file, wouldn't logging the failure and sticking with the previous version of the file be preferable?
That's exactly the point (ie just prior to distribution) where a simple sanity check should have been run and the config replacement/update pipeline stopped on failure. When they introduced the 200 entry limit memory optimised feature loader it should have been a no-brainer to insert that sanity check in the config production pipeline.
Yea, Rust is safe but it’s not magic. However Nginx doesn’t panic on malformed config. It exits with hopefully a helpful error code and message. The question is then could the cloudflare code have exited cleanly in a way that made recovery easier instead of just straight panicking.
Would expect with a message meet that criteria of exiting with a more helpful error message? From the postmortem it seems to me like they just didn’t know it even was panicing
One feature failing like this should probably log the error and fail closed. It shouldn't take down everything else in your big proxy that sits in front of your entire business.
Exactly! Sometimes exploding is simply the least bad option, and is an entirely sensible approach.
Falling back to a generic base configuration in the presence of an incoming invalid config file would probably be a sensible thing to do.
I'm not a fan of rust, but I don't think that is the only takeaway. All systems have assumptions about their input and if the assumption is violated, it has to be caught somewhere. It seems like it was caught too deep in the system.
Maybe the validation code should've handled the larger size, but also the db query produced something invalid. That shouldn't have ever happened in the first place.
> It seems like it was caught too deep in the system.
Agreed, that's also my takeaway.
I don't see the problem being "lazy programmers shouldn't have called .unwrap()". That's reductive. This is a complex system and complex system failures aren't monocausal.
The function in question could have returned a smarter error rather than panicking, but what then? An invariant was violated, and maybe this system, at this layer, isn't equipped to take any reasonable action in response to that invariant violation and dying _is_ the correct thing to do.
But maybe it could take smarter action. Maybe it could be restarted into a known good state. Maybe this service could be supervised by another system that would have propagated its failure back to the source of the problem, alerting operators that a file was being generated in such a way that violated consumer invariants. Basically, I'm describing a more Erlang model of failure.
Regardless, a system like this should be able to tolerate (or at least correctly propagate) a panic in response to an invariant violation.
Safe things should be easy, dangerous things should be hard.
This .unwrap() sounds too easy for what it does, certainly much easier than having an entire try..catch block with an explicit panic. Full disclosure: I don't actually know Rust.
I don't think 'unwrap' is inherently the problem.
Any project has to reason about what sort of errors can be tolerated gracefully and which cannot. Unwrap is reasonable in scenarios you expect to never be reached, because otherwise your code will be full of all sorts of possible permutations and paths that are harder to reason about and may cascade into extremely nuanced or subtle errors.
Rust also has a version of unwrap called "expect" where you provide a string that logs why the unwrap occurred. It's similar, but for pieces of code that are crucial it could be a good idea to require all 'unwraps' to instead be 'expects' so that people at least are forced to write down a reason why they believe the unwrap can never be reached.
Not panicking code is tedious to write. It is not realistic to expect everything to be non panic. There is a reason that panicking exists in the first place.
Them calling unwrap on a limit check is the real issue imo. Everything that takes in external input should assume it is bad input and should be fuzz tested imo.
In the end, what is the point of having a limit check if you are just unwrapping on it
> Not panicking code is tedious to write.
Using the question mark operator [1] and even adding in some anyhow::context goes a long way to being able to fail fast and return an Err rather then panicking.
Sure you need to handle Results all the way up the stack but at least it forces you to think about how those nested parts of your app will fail as you travel back up the stack.
[1]: https://doc.rust-lang.org/rust-by-example/std/result/questio...
git blame on .unwrap() line
As a gopher I never understand why is there so many unwraps in an average rust code.
Average Go code has much less panics than Rust has unwraps, which are functionally equivalent.
The average golang code segfaults by design.
Because Go silently gives you zero/null instead
Idiomatically, it gives you `err` and you do `if err != nil {return err}`. While in rust you mostly do `.unwrap` and panic.
It's not in the type system, but it's idiomatic
which mean an unexpected behavior could go unnoticed for a long time.
I'd prefer a loud crash over that.
Well look at the failure modes in the original article.
In the original PHP code, all worked, only it didn't properly check for bots.
The new Rust code did a loud crash and took off half of the internet.
Swift has implicit unwrap (!), and explicit unwrap (?).
I don't like to use implicit unwrap. Even things that are guaranteed to be there, I treat as explicit (For example, (self.view?.isEnabled ?? false), in a view controller, instead of self.view.isEnabled).
I always redefine @IBOutlets from:
to: I'm kind of a "belt & suspenders" type of guy.So what happens if it ends up being nil? How does your app react?
In this particular case, I would rather crash. It’s easier to spot in a crash report and you get a nice stack trace.
Silent failure is ultimately terrible for users.
Note: for the things I control I try to very explicitly model state in such a way as I never need to force unwrap at all. But for things beyond my control like this situation, I would rather end the program than continue with a state of the world I don’t understand.
Yeah @IBOutlets are generally the one thing that are allowed to be implicitly-unwrapped optionals. They go along with using storyboards & xibs files with Interface Builder. I agree that you really should just crash if you are attempting to access one and it is nil. Either you have done something completely incorrect with regards to initializing and accessing parts of your UI and want to catch that in development, or something has gone horribly, horribly, horribly with UIKit/AppKit and storyboard/xib files are not being loaded properly by the system.
> … you really should just crash if …
See my above/below comment.
A good tool for catching stuff during development, is the humble assert()[0]. We can use precondition()[1], to do the same thing, in ship code.
The main thing is, is to remain in control, as much as possible. Rather than let the PC leave the stack frame, throw the error immediately when it happens.
[0] https://docs.swift.org/swift-book/documentation/the-swift-pr...
[1] https://docs.swift.org/swift-book/documentation/the-swift-pr...
> Silent failure is ultimately terrible for users.
Agreed.
Unfortunately, crashes in iOS are “silent failures,” and are a loss of control.
What this practice does, is give me the option to handle the failure “noisily,” and in a controlled manner; even if just emitting a log entry, before calling a system failure. That can be quite helpful, in threading. Also, it gives me the option to have a valid value applied, if there’s a structural failure.
But the main reason that I do that with @IBOutlets, is that it forces me to acknowledge, throughout the rest of the code, that it’s an optional. I could always treat implicit optionals as if they were explicit, anyway. This just forces me to.
I have a bunch of practices that folks can laugh at, but my stuff works pretty effectively, and I sleep well.
It rang more as "A/B deployments are pointless if you can't tell if a downstream failure is related." To me.
That's such a bad take after reading the article. If you're going to write a system that preallocates and is based on hard assumptions about max size - the panic/unwrap approach is reasonable.
The config bug reaching prod without this being caught and pinpointed immediately is the strange part.
It's reasonable when testing protocols exercise the panic scenario. This is the problem with punting on error recovery. Nobody checks faults that propagate across domains of responsibility.
I agree there's no way to soft-error this, though "truncate and raise an alert" is arguably the better pattern.
Exactly. The newbie mistake in SQL is also way worse than this. But the whole design is also bad. Clearly implementing things at the wrong place.
And, it took like over an hour between the problem started til my sites went down. That is just crazy.
This is a bummer. The unwrap()'ing function already returned a result and should have just propagated the error. Presumably the caller could have handled more sensibly than just panic'ing.
Some languages and style guides simply forbid throwing exceptions without catching / proper recovery. Google C++ bans exceptions and the main mechanism for propogating errors is `absl::Status` which the caller has to check. Not familiar with Rust but it seems unwrap is such a thing that would be banned.
> Not familiar with Rust but it seems unwrap is such a thing that would be banned.
Panics aren't exceptions, any "panic" in Rust can be thought of as an abort of the process (Rust binaries have the explicit option to implement panics as aborts). Companies like Dropbox do exactly this in their similar Rust-based systems, so it wouldn't surprise me if Cloudflare does the same.
"Banning exceptions" wouldn't have done anything here, what you're looking for is "banning partial functions" (in the Haskell sense).
Yeah I know but isn't unwrap basically a trivial way to: (1) give up catching the exception/error (the E part in `Result<T, E>`) that the callee throws; and also (2) escalate it to the point that nothing can catch it. It has such a benign name.
Unwrap is used in places where in C++ you would just have undefined behavior. It wouldn't make any more sense to blanket ban it than it would ban ever dereferencing a pointer just in case its null - even if you just checked that it wasn't null.
Rust's Result is the same thing as C++'s std::expected. How is calling std::expected::value undefined behaviour?
Rust's foo: Option<&T> is rust's rough equivalent to C++'s const T* foo. The C++ *foo is equivalent to the rust unsafe{ *foo.unwrap_unchecked() }, or in safe code *foo.unwrap() (which changes the undefined behavior to a panic).
Rust's unwrap isn't the same as std::expected::value. The former panics - i.e. either aborts the program or unwinds depending on context and is generally not meant to be handled. The latter just throws an exception that is generally expected to be handled. Panics and exceptions use similar machinery (at least they can depending on compiler options) but they are not equivalent - for example nested panics in destructors always abort the program.
In code that isn't meant to crash `unwind` should be treated as a sign saying that "I'm promising that this will never happen", but just like in C++ where you promise that pointers you deference are valid and signed integers you add don't overflow making promises like that is a necessary part of productive programming.
Tangential but funnily enough calling std::expected::error is ub if there is no error :D
There are even lints for this but people get impatient and just override them or fight for them to no longer be the default.
As usual: people problem, not a tech problem. In the last years a lot of strides have been made. But people will be people.
Linting is not good enough. The compiler should refuse to compile the code without it marked with an explicit annotation. Too much Rust code is panic happy since using casual use of `unwrap` is perma-etched into everyone's minds by the amount of demo code out there using unwrap.
!#[deny(unwrap_used)] at the top of the file would take care of it.
I completely agree. But IMO Rust would have not gained traction if it was as strict. It would be branded as an academic toy language.
But now after we are past that and it has a lot of mind share, I'd say it's time to start tightening the bolts.
and people make mistake
at some point machine would be better in coding because well machine code is machine instruction task
same like chess, engine is better than human grandmaster because its solvable math field
coding is no different
> same like chess, engine is better than human grandmaster because its solvable math field
Might be worth noting that your description of chess is slightly incorrect. Chess technically isn't solved in the sense that the optimal move is known for any arbitrary position is known; it's just that chess engines are using what amounts to a fancy brute force for most of the game and the combination of hardware and search algorithm produces a better result than the human brain does. As such, chess engines are still capable of making mistakes, even if actually exploiting them is a challenge.
No ?????? because these thing called BEST MOVE and BAD MOVE there in chess
"chess engines are still capable of making mistakes", I'm sorry no
inaccurate yes but not mistake
> because these thing called BEST MOVE and BAD MOVE there in chess
The thing is that there is no known general objective criteria for "best" and "bad" moves. The best we have so far is based on engine evaluations, but as I said before that is because chess engines are better at searching the board's state space than humans, not because chess engines have solved chess in the mathematical sense. Engines are quite capable of misevaluating positions, as demonstrated quite well by the Top Chess Engine Championship [0] where one engine thinks it made a good move while the other thinks that move is bad, and this is especially the case when resources are limited.
The closest we are to solving chess are via tablebases, which are far from covering the entire state space and are basically as much of an exemplar of pure brute force as you can get.
> "chess engines are still capable of making mistakes", I'm sorry no
If you think chess engines are infalliable, then why does the Top Chess Engine Championship exist? Surely if chess engines could not make mistakes they would always agree on a position's evaluation and what move should be made, and therefore such an exercise would be pointless?
> inaccurate yes but not mistake
From the perspective to attaining perfect play an inaccuracy is a mistake.
[0]: https://en.wikipedia.org/wiki/Top_Chess_Engine_Championship
"The thing is that there is no known general objective criteria for "best" and "bad" moves."
are you playing chess or not?????? if you playing chess then its oblivious how to differentiate bad move and best move
Yes it is objective, these thing called best move not without reason
"If you think chess engines are infalliable, then why does the Top Chess Engine Championship exist?"
to create better chess engine like what do even talking about here????, are you saying just because there are older bad engine that mean this thing is pointless ????
if you playing chess up to a decent level 1700+ (like me), you know that these argument its wrong and I assure you to learn chess to a decent level
up until that point that you know high level chess is brute force games and therefore solvable math
> if you playing chess then its oblivious how to differentiate bad move and best move
The key words in what I said are "general" and "objective". Yes, it's possible to determine "good" or "bad" moves in specific positions. There's no known method to determine "good" or "bad" moves in arbitrary positions, as would be required for chess to be considered strongly solved.
Furthermore, if it's "obvious" how to differentiate good and bad moves then we should never see engines blundering, right?
So (for example) how do you explain this game between Stockfish and Leela where Stockfish blunders a seemingly winning position [0]? After 37... Rdd8 both Stockfish and Leela think white is clearly winning (Stockfish's evaluation is +4.00, while Leela's evaluation is +3.81), but after 38. Nxb5 Leela's evaluation plummets to +0.34 while Stockfish's evaluation remains at +4.00. In the end, it turns out Leela was correct after 40... Rxc6 Stockfish's evaluation also drops from +4.28 to 0.00 as it realizes that Leela has a forced stalemate.
Or this game also between Stockfish and Leela where Leela blunders into a forced mating sequence and doesn't even realize it for a few moves [1]?
Engines will presumably always play what they think is the "best" move, but clearly sometimes this "best" move is wrong. Evidently, this means differentiating "good" and "bad" moves is not always obvious.
> Yes it is objective, these thing called best move not without reason
If it's objective, then why is it possible for engines to disagree on whether a move is good or bad, as they do in the above example and others?
> to create better chess engine like what do even talking about here????
The ability to create better chess engines necessarily implies that chess engines can and do make mistakes, contrary to what you asserted.
> are you saying just because there are older bad engine that mean this thing is pointless ????
No. What I'm saying is that your explanation for why chess engines are better than humans is wrong. Chess engines are not better than humans because they have solved chess in the mathematical sense; chess engines are better than humans because they search the state space faster and more efficiently than humans (at least until you reach 7 pieces on the board).
> up until that point that you know high level chess is brute force games and therefore solvable math
"Solvable" and "solved" are two very different things. Chess is solvable, in theory. Chess is very far from being solved.
[0]: https://www.chess.com/computer-chess-championship#event=309&...
[1]: https://www.chess.com/computer-chess-championship#event=309&...
Which is something I will bookmark for the usual Rust doesn't do exceptions discussions, except it kind of does even if called differently.
I'm not completely sure I agree. I mean, I do agree about the .unwrap() culture being a bug trap. But I don't think this example qualifies.
The root cause here was that a file was mildly corrupt (with duplicate entries, I guess). And there was a validation check elsewhere that said "THIS FILE IS TOO BIG".
But if that's a validation failure, well, failing is correct? What wasn't correct was that the failure reached production. What should have happened is that the validation should have been a unified thing and whatever generated the file should have flagged it before it entered production.
And that's not an issue with function return value API management. The software that should have bailed was somewhere else entirely, and even there an unwrap explosion (in a smoke test or pre-release pass or whatever) would have been fine.
It sounds to me like there was validation, but the system wasn't designed for the validation to ever fail - at which point crashing is the only remaining option. You've essentially turned it into an assertion error rather than a parsing/validation error.
Ideally every validation should have a well-defined failure path. In the case of a config file rotation, validation failure of the new config could mean keeping the old config and logging a high-priority error message. In the case of malformed user-provided data, it might mean dropping the request and maybe logging it for security analysis reasons. In the case of "pi suddenly equals 4" checks the most logical approach might be to intentionally crash, as there's obviously something seriously wrong and application state has corrupted in such a way that any attempt to continue is only going to make things worse.
But in all cases there's a reason behind the post-validation-failure behavior. At a certain point leaving it up to "whatever happens on .unwrap() failure" isn't good enough anymore.
I wonder what happens if they handle it gracefully? sounds like performance degradation (better than reliability degradation!).
Also wonder with a sharded system why are they not slow rolling out changes and monitoring?
In addition, it looks like this system wasn't on any kind of 1%/10%/50%/100% rollout gating. Such a rollout would trivially have shown the poison input killing tasks.
To me it reads like there was a gradual rollout of the faulty software responsible for generating the config files, but those files are generated on approximately one machine, then propogated across the whole network every 5 minutes.
> Bad data was only generated if the query ran on a part of the cluster which had been updated. As a result, every five minutes there was a chance of either a good or a bad set of configuration files being generated and rapidly propagated across the network.
Not a DBA, how do you do DB permission rollout gating?
It looks like changing the permissions triggered creation of a new feature file, and it was ingestion of that file leading to blowing a size limit that crashed the systems.
The file should be versioned and rollout of new versions should be staged.
(There is definitely a trade-off; often times in the security critical path, you want to go as fast as possible because changes may be blocking a malicious actor. But if you move too fast, you break things. Here, they had a potential poison input in the pathway for synchronizing this state and Murphy's Law suggests it was going to break eventually, so the question becomes "How much damage can we tolerate when it does?")
> It looks like changing the permissions triggered creation of a new feature file, and it was ingestion of that file leading to blowing a size limit that crashed the systems.
That feature file is generated every 5 minutes at all times; the change to permissions was rolled out gradually over the clickhouse cluster, and whether a bad version of that file was generated depended on whether the part of the cluster that had the bad permissions generated the file.
> This is textbook "parse, don't validate" anti-pattern.
How so? “Parse, don’t validate” implies converting input into typed values that prevent representation of invalid state. But the parsing still needs to be done correctly. An unchecked unwrap really has nothing to do with this.
GP completely misunderstands “parse, don’t validate” and also calls it an anti-pattern. GP clearly has no idea what this is.
if you make it easy to be lazy and panic vs properly handling the error, you've designed a poor language
At Facebook they name certain "escape hatch" functions in a way that inescapably make them look like a GIANT EYESORE. Stuff like DANGEROUSLY_CAST_THIS_TO_THAT, or INVOKE_SUPER_EXPENSIVE_ACTION_SEE_YOU_ON_CODE_REVIEW. This really drives home the point that such things must not be used except in rare extraordinary cases.
If unwrap() were named UNWRAP_OR_PANIC(), it would be used much less glibly. Even more, I wish there existed a super strict mode when all places that can panic are treated as compile-time errors, except those specifically wrapped in some may_panic_intentionally!() or similar.
> make them look like a GIANT EYESORE
React.__SECRET_INTERNALS_DO_NOT_USE_OR_YOU_WILL_BE_FIRED comes to mind. I did have to reach to this before, but it certainly works for keeping this out of example code and other things like reading other implementations without the danger being very apparent.
At some point it was renamed to __CLIENT_INTERNALS_DO_NOT_USE_OR_WARN_USERS_THEY_CANNOT_UPGRADE which is much less fun.
> it certainly works
Not for this guy:
https://github.com/reactjs/react.dev/issues/3896
right and if the language designers named it UNWRAP_OR_PANIC() then people would rightfully be asking why on earth we can't just use a try-catch around code and have an easier life
But a panic can be caught and handled safely (e.g. via std:: panic tools). I'd say that this is the correct use case for exceptions (ask Martin Fowler, of all people).
There is already a try/catch around that code, which produces the Result type, which you can presumptuously .unwrap() without checking if it contains an error.
Instead, one should use the question mark operator, that immediately returns the error from the current function if a Result is an error. This is exactly similar to rethrowing an exception, but only requires typing one character, the "?".
Probably not, since errors as values are way better than exceptions.
How so? An exception is a value that's given the closest, conceptually appropriate, point that was decided to handle the value, allowing you to keep your "happy path" as clean code, and your "exceptional circumstances" path at the level of abstraction that makes sense.
It's way less book-keeping with exceptions, since you, intentionally, don't have to write code for that exceptional behavior, except where it makes sense to. The return by value method, necessarily, implements the same behavior, where handling is bubbled up to the conceptually appropriate place, through returns, but with much more typing involved. Care is required for either, since not properly bubbling up an exception can happen in either case (no re-raise for exceptions, no return after handling for return).
There are many many pages of text discussing this topic, but having programmed in both styles, exceptions make it too easy for programmer to simply ignore them. Errors as values force you to explicitly handle it there, or toss it up the stack. Maybe some other languages have better exception handling but in Python it’s god awful. In big projects you can basically never know when or how something can fail.
I would claim the opposite. If you don't catch an exception, you'll get a halt.
With return values, you can trivially ignore an exception.
In the wild, I've seen far more ignoring return errors, because of the mechanical burden of having type handling at every function call.This is backed by decades of writing libraries. I've tried to implement libraries without exceptions, and was my admittedly cargo-cult preference long ago, but ignoring errors was so prevalent among the users of all the libraries that I now always include a "raise" type boolean that defaults to True for any exception that returns an error value, to force exceptions, and their handling, as default behavior.
> In big projects you can basically never know when or how something can fail.
How is this fundamentally different than return value? Looking at a high level function, you can't know how it will fail, you just know it did fail, from the error being bubbled up through the returns. The only difference is the mechanism for bubbling up the error.
Maybe some water is required for this flame war. ;)
I can agree to disagree :)
Exception is hidden control flow, where as error values are not.
That is the main reason why zig doesn’t have exceptions.
I'd categorize them more as "event handlers" than "hidden". You can't know where the execution will go at a lower level, but that's the entire point: you don't care. You put the handlers at the points where you care.
Correction: unchecked exceptions are hidden control flow. Checked exceptions are quite visible, and I think that more languages should use them as a result.
...and you can? try-catch is usually less ergonomic than the various ways you can inspect a Result.
vs Or better yet, compare the problematic cases where the error isn't handled: vs In the former (the try-catch version that doesn't try or catch), the lack of handling is silent. It might be fine! You might just depend on your caller using `try`. In the latter, the compiler forces you to use UNWRAP_OR_PANIC (or, in reality, just unwrap) or `data` won't be the expected type and you will quickly get a compile failure.What I suspect you mean, because it's a better argument, is:
which is fair, although how often is it really the right thing to let all the errors from 4 independent sources flow together and then get picked apart after the fact by inspecting `e`? It's an easier life, but it's also one where subtle problems constantly creep in without the compiler having any visibility into them at all.Unwrap isn't a synonym for laziness, it's just like an assertion, when you do unwrap() you're saying the Result should NEVER fail, and if does, it should abort the whole process. What was wrong was the developer assumption, not the use of unwrap.
It also makes it very obvious in the code, something very dangerous is happening here. As a code reviewer you should see an unwrap() and have alarm bells going off. While in other languages, critical errors are a lot more hidden.
I hate that it's a method. That can get lost in a method chain easily enough during a code review.
A function or a keyword would interrupt that and make it less tempting
Well, you can request Clippy to tell you about them. I do that in my hobby projects.
> What was wrong was the developer assumption, not the use of unwrap.
How many times can you truly prove that an `unwrap()` is correct and that you also need that performance edge?
Ignoring the performance aspect that often comes from a hat-trick, to prove such a thing you need to be wary of the inner workings of a call giving you a `Return`. That knowledge is only valid at the time of writing your `unwrap()`, but won't necessarily hold later.
Also, aren't you implicitly forcing whoever changes the function to check for every smartass dev that decided to `unwrap` at their callsite? That's bonkers.
I doubt that this unwrap was added for performance reasons; I suspect it was rather added because the developer temporarily didn't want to deal with what they thought was an unlikely error case while they were working on something else; and no other system recognized that the unwrap was left in and flagged it before it was deployed on production servers.
If I were Cloudflare I would immediately audit the codebase for all uses of unwrap (or similar rust panic idioms like expect), ensure that they are either removed or clearly documented as to why it's worth crashing the program there, and then add a linter to their CI system that will fire if anyone tries to check in a new commit with unwrap in it.
Panics are for unexpected error conditions, like your caller passed you garbage. Results are for expected errors, like your caller passed you something but it's your job to tell if it's garbage.
So the point of unwrap() is not to prove anything. Like an assertion it indicates a precondition of the function that the implementer cannot uphold. That's not to say unwrap() can't be used incorrectly. Just that it's a valid thing to do in your code.
Note that none of this is about performance.
> when you do unwrap() you're saying the Result should NEVER fail
Returning a Result by definition means the method can fail.
> Returning a Result by definition means the method can fail.
No more than returning an int by definition means the method can return -2.
> No more than returning an int by definition means the method can return -2.
What? Returning an int does in fact mean that the method can return -2. I have no idea what your argument is with this, because you seem to be disagreeing with the person while actually agreeing with them.
What? Results have a limited number of possible error states that are well defined.
Some call points to a function that returns a Result will never return an Error.
Some call points to a function that returns an int will never return -2.
Sometimes you know things the type system does not know.
The difference is functions which return Result have explicitly chosen to return a Result because they can fail. Sure, it might not fail in the current implementation and/or configuration, but that could change later and you might not know until it causes problems. The type system is there to help you - why ignore it?
Because it would be a huge hassle to go into that library and write an alternate version that doesn't return a Result. So you're stuck with the type system being wrong in some way. You can add error-handling code upfront but it will be dead code at that point in time, which is also not good.
Yeah, I think I expressed wrongly here. A more correct version would be: "when you do unwrap() you're saying that an error on this particular path shouldn't be recoverable and we should fail-safe."
Division can fail, but I can write a program where I'm sure division will not fail.
It's a little subtler than this. You want it to be easy to not handle an error while developing, so you can focus on getting the core logic correct before error-handling; but you want it to be hard to deploy or release the software without fully handling these checks. Some kind of debug vs release mode with different lints seems like a reasonable approach.
So… basically every language ever?
Except maybe Haskell.
And Gleam
In Rust, `.unwrap()` is nine characters, whereas propagating the Result via `?` is one.
All languages with few exceptions have these kinds of escape hatches like unwrap
https://en.wikipedia.org/wiki/Crash-only_software
Works when you have the Erlang system that does graceful handing for you: reporting, restarting.
it's usually because of fail fast and fail hard, in theory critical bugs will be caught in dev/test
You write so much rust you causally apply unwrap now to everything?
Rust compiler is a god of sorts, or at least a law of nature haha
Way to comment and go instantly off topic
> This showed up to Internet users trying to access our customers' sites as an error page indicating a failure within Cloudflare's network.
As a visitor to random web pages, I definitely appreciated this—much better than their completely false “checking the security of your connection” message.
> The issue was not caused, directly or indirectly, by a cyber attack or malicious activity of any kind. Instead, it was triggered by a change to one of our database systems' permissions
Also appreciate the honesty here.
> On 18 November 2025 at 11:20 UTC (all times in this blog are UTC), Cloudflare's network began experiencing significant failures to deliver core network traffic. […]
> Core traffic was largely flowing as normal by 14:30. We worked over the next few hours to mitigate increased load on various parts of our network as traffic rushed back online. As of 17:06 all systems at Cloudflare were functioning as normal.
Why did this take so long to resolve? I read through the entire article, and I understand why the outage happened, but when most of the network goes down, why wasn't the first step to revert any recent configuration changes, even ones that seem unrelated to the outage? (Or did I just misread something and this was explained somewhere?)
Of course, the correct solution is always obvious in retrospect, and it's impressive that it only took 7 minutes between the start of the outage and the incident being investigated, but it taking a further 4 hours to resolve the problem and 8 hours total for everything to be back to normal isn't great.
> much better than their completely false “checking the security of your connection” message
The exact wording (which I can easily find, because a good chunk of the internet gives it to me, because I’m on Indian broadband):
> example.com needs to review the security of your connection before proceeding.
It bothers me how this bald-faced lie of a wording has persisted.
(The “Verify you are human by completing the action below.” / “Verify you are human” checkbox is also pretty false, as ticking the box in no way verifies you are human, but that feels slightly less disingenuous.)
Because we initially thought it was an attack. And then when we figured it out we didn’t have a way to insert a good file into the queue. And then we needed to reboot processes on (a lot) of machines worldwide to get them to flush their bad files.
Thanks for the explanation! This definitely reminds me of CrowdStrike outages last year:
- A product depends on frequent configuration updates to defend against attackers.
- A bad data file is pushed into production.
- The system is unable to easily/automatically recover from bad data files.
(The CrowdStrike outages were quite a bit worse though, since it took down the entire computer and remediation required manual intervention on thousands of desktops, whereas parts of Cloudflare were still usable throughout the outage and the issue was 100% resolved in a few hours)
Richard Cook #18 (and #10) strikes again!
https://how.complexsystems.fail/#18
It'd be fun to read more about how you all procedurally respond to this (but maybe this is just a fixation of mine lately). Like are you tabletopping this scenario, are teams building out runbooks for how to quickly resolve this, what's the balancing test for "this needs a functional change to how our distributed systems work" vs. "instead of layering additional complexity on, we should just have a process for quickly and maybe even speculatively restoring this part of the system to a known good state in an outage".
Why was Warp in London disabled temporarily. No mention of that change was discussed in the RCA despite it being called out in an update.
For London customers this made the impact more severe temporarily.
Probably because it was the London team that was actively investigating the incident and initially came to the conclusion that it may be a DDoS while being unable to authenticate to their own systems.
Given the time of the outage that makes sense, they’d mostly be within their work day time (if such a thing apples to us anymore).
Question from a casual bystander, why not have a virtual/staging mini node that receives these feature file changes first and catches errors to veto full production push?
Or you do have something like this but the specific db permission change in this context only failed in production
I think the reasoning behind this is because of the nature of the file being pushed - from the post mortem:
"This feature file is refreshed every few minutes and published to our entire network and allows us to react to variations in traffic flows across the Internet. It allows us to react to new types of bots and new bot attacks. So it’s critical that it is rolled out frequently and rapidly as bad actors change their tactics quickly."
In this case, the file fails quickly. A pretest that consists of just attempting to load the file would have caught it. Minutes is more than enough time to perform such a check.
Thx for the explanation!
Side thought as we're working on 100% onchain systems (for digital assets security, different goals):
Public chains (e.g. EVMs) can be a tamper‑evident gate that only promotes a new config artifact if (a) a delay or multi‑sig review has elapsed, and (b) a succinct proof shows the artifact satisfies safety invariants like ≤200 features, deduped, schema X, etc.
That could have blocked propagation of the oversized file long before it reached the edge :)
Just asking out of curiosity, but roughly how many staff would've been involved in some way in sorting out the issue? Either outside regular hours or redirected from their planned work?
Yeah, I can imagine that this insertion was some high-pressure job.
As always, kudos for releasing a post mortem in less than 24 hours after the outage, very few tech organisations are capable of doing this.
I'm curious about how their internal policies work such that they are allowed to publish a post mortem this quickly, and with this much transparency.
Any other large-ish company, there would be layers of "stakeholders" that will slow this process down. They will almost always never allow code to be published.
Well… we have a culture of transparency we take seriously. I spent 3 years in law school that many times over my career have seemed like wastes but days like today prove useful. I was in the triage video bridge call nearly the whole time. Spent some time after we got things under control talking to customers. Then went home. I’m currently in Lisbon at our EUHQ. I texted John Graham-Cumming, our former CTO and current Board member whose clarity of writing I’ve always admired. He came over. Brought his son (“to show that work isn’t always fun”). Our Chief Legal Officer (Doug) happened to be in town. He came over too. The team had put together a technical doc with all the details. A tick-tock of what had happened and when. I locked myself on a balcony and started writing the intro and conclusion in my trusty BBEdit text editor. John started working on the technical middle. Doug provided edits here and there on places we weren’t clear. At some point John ordered sushi but from a place with limited delivery selection options, and I’m allergic to shellfish, so I ordered a burrito. The team continued to flesh out what happened. As we’d write we’d discover questions: how could a database permission change impact query results? Why were we making a permission change in the first place? We asked in the Google Doc. Answers came back. A few hours ago we declared it done. I read it top-to-bottom out loud for Doug, John, and John’s son. None of us were happy — we were embarrassed by what had happened — but we declared it true and accurate. I sent a draft to Michelle, who’s in SF. The technical teams gave it a once over. Our social media team staged it to our blog. I texted John to see if he wanted to post it to HN. He didn’t reply after a few minutes so I did. That was the process.
> I texted John to see if he wanted to post it to HN. He didn’t reply after a few minutes so I did
Damn corporate karma farming is ruthless, only a couple minute SLA before taking ownership of the karma. I guess I'm not built for this big business SLA.
You call this transparency, but fail to answer the most important questions: what was in the burrito? Was it good? Would you recommend?
Chicken burrito from Coyo Taco in Lisbon. I am not proud of this. It’s worse than ordering from Chipotle. But there are no Chipotle’s in Lisbon… yet.
Appreciate the extra transparency on the process.
A very human and authentic response. Love to see it.
Fantastic for recruiting, too.
> He didn’t reply after a few minutes so I did
I'd consider applying based on this alone
How do you guys handle redaction? I'm sure even when trusted individuals are in charge of authoring, there's still a potential of accidental leakage which would probably be best mitigated by a team specifically looking for any slip ups.
Thanks for the insight.
A postmortem postmortem, I love it. Transparency to the power of 2.
that's very cool, thanks
I mean the CEO posted the post-mortem so there aren't that many layers of stakeholders above. For other post-mortems by engineers, Matthew once said that the engineering team is running the blog and that he wouldn't event know how to veto even if he wanted [0]
[0] https://news.ycombinator.com/item?id=45588305
The person who posted both this blog article and the hacker news post, is Matthew Prince, one of highly technical billionaire founders of cloudflare. I'm sure if he wants something to happen, it happens.
From what I've observed, it depends on whether you're an "engineering company" or not.
And a well-written one at that. Compared to the AWS port-mortem this could be literature.
Except it fails to document anything about the actions they made to Warp in London during the resolution.
Plenty are capable. Most don't bother.
I’m surprised by how quickly they put together a post mortem, let alone how quickly they were willing to publish it to the world.
Meanwhile, on X, white nationalist supremacists / Groypers / America First America Only types are spreading a conspiracy theory that Cloudflare’s outage was caused by having too many Indians employees and H1Bs. Example:
https://xcancel.com/DiggingInTheDi1/status/19908503102026673...
This is far from the only one such post, but what’s amusing / horrifying is that this belief is getting millions of likes across all these posts.
Why give this sort of content more visibility/reach?
I'm sure that's not your intent, so I hope my comment gives you an opportunity to reflect on the effects of syndicating such stupidity, no matter what platform it comes from.
Mainly to make others aware of what’s happening in the context of this Cloudflare outage. Sure I can avoid giving it visibility/reach but it’s growing and proliferating on its own, and I think ignoring it isn’t going to stop it so I am hoping awareness will help. I’ve noticed a huge rise in open racism against Chinese and Indian and workers of other origin, even when they’re here on a legal visa that we have chosen as a nation to grant for our own benefit.
The legislation that MTG (Marjorie Taylor Green) just proposed a few days ago to ban H1B entirely, and the calls to ban other visa types, is going to have a big negative impact on the tech industry and American innovation in general. The social media stupidity is online but it gives momentum to the actual real life legislation and other actions the administration might take. Many congress people are seeing the online sentiment and changing their positions in response, unfortunately.
Fair points; there's certainly a balance to be struck between raising awareness and amplifying, and I admittedly have no idea where that line is.
The unwrap: not great, but understandable. Better to silently run with a partial config while paging oncall on some other channel, but that's a lot of engineering for a case that apparently is supposed to be "can't happen".
The lack of canary: cause for concern, but I more or less believe Cloudflare when they say this is unavoidable given the use case. Good reason to be extra careful though, which in some ways they weren't.
The slowness to root cause: sheer bad luck, with the status page down and Azure's DDoS yesterday all over the news.
The broken SQL: this is the one that I'd be up in arms about if I worked for Cloudflare. For a system with the power to roll out config to ~all of prod at once while bypassing a lot of the usual change tracking, having this escape testing and review is a major miss.
Share the same opinion, as others pointed out, the status page down probably caused by bots checking it.
> That feature file, in turn, doubled in size. The larger-than-expected feature file was then propagated to all the machines that make up our network.
> The software running on these machines to route traffic across our network reads this feature file to keep our Bot Management system up to date with ever changing threats. The software had a limit on the size of the feature file that was below its doubled size. That caused the software to fail.
I'm no FAANG 10x engineer, and I appreciate things can be obvious in hindsight, but I'm somewhat surprised that engineering at the level of Cloudflare does not:
1. Push out files A/B to ensure the old file is not removed.
2. Handle the failure of loading the file (for whatever reason) by automatically reloading the old file instead and logging the error.
This seems like pretty basic SRE stuff.
It reads a lot like the Crowdstrike SNAFU. Machine-generated configuration file b0rks-up the software that consumes it.
The "...was then propagated to all the machines that make up our network..." followed by "....caused the software to fail." screams for a phased rollout / rollback methodology. I get that "...it’s critical that it is rolled out frequently and rapidly as bad actors change their tactics quickly" but today's outage highlights that rapid deployment isn't all upside.
The remediation section doesn't give me any sense that phased deployment, acceptance testing, and rapid rollback are part of the planned remediation strategy.
At my employer, we have a small script that automatically checks such generated config files. It does a diff between the old and the new version, and if the diff size exceeds a threshold (either total or relative to the size of the old file), it refuses to do the update, and opens a ticket for a human to look over it.
It has somewhat regularly saved us from disaster in the past.
I don't think this system is best thought of as "deployment" in the sense of CI/CD; it's a control channel for a distributed bot detection system that (apparently) happens to be actuated by published config files (it has a consul-template vibe to it, though I don't know if that's what it is).
Code and Config should be treated similarly. If you would use a ring based rollout, canaries, etc for safely changing your code, then any config that can have the same impact must also use safe rollout techniques.
That's why I likened it Crowdstrike. It's a signature database that blew up the consumer of said database. (You probably caught my post mid-edit, too. You may be replying to the snarky paragraph I felt better of and removed.)
Edit: Similar to Crowdstrike, the bot detector should have fallen-back to its last-known-good signature database after panicking, instead of just continuing to panic.
That’s correct.
Is it actually consul-template? (I have post-consul-template stress disorder).
I'd love to hear any commentary on Consul if anyone else has it.
I think Consul is great, for what it's worth; we were just abusing it.
https://fly.io/blog/a-foolish-consistency/
https://fly.io/blog/corrosion/
Did you know: PCTSD affects more than 2 in 5 engineers.
I'm amazed that they are not using any simulator of some sort and pushing changes directly to production.
I’m fairly certain it will be after they read this thread. It doesn’t feel like they don’t want, or are incapable of improving?
> work has already begun on how we will harden them against failures like this in the future. In particular we are:
> Hardening ingestion of Cloudflare-generated configuration files in the same way we would for user-generated input
> Enabling more global kill switches for features
> Eliminating the ability for core dumps or other error reports to overwhelm system resources
> Reviewing failure modes for error conditions across all core proxy modules
Absent from this list are canary deployments and incremental or wave-based deployment of configuration files (which are often as dangerous as code changes) across fault isolation boundaries -- assuming CloudFlare has such boundaries at all. How are they going to contain the blast radius in the future?
This is something the industry was supposed to learn from the CrowdStrike incident last year, but it's clear that we still have a long way to go.
Also, enabling global anything (i.e., "enabling global kill switches for features") sounds like an incredibly risky idea. One can imagine a bug in a global switch that transforms disabling a feature into disabling an entire system.
They require the bot management config to update and propagate quickly in order to respond to attacks - but this seems like a case where updating a since instance first would have seen the panic and stopped the deploy.
I wonder why clickhouse is used to store the feature flags here, as it has it's own duplication footguns[0] which could have also easily lead to a query blowing up 2/3x in size. oltp/sqlite seems more suited, but i'm sure they have their reasons
[0] https://clickhouse.com/docs/guides/developer/deduplication
I don't think sqlite would come close to their requirements for permissions or resilience, to name a couple. It's not the solution for every database issue.
Also, the link you provided is for eventual deduplication at the storage layer, not deduplication at query time.
I think the idea is to ship the sqlite database around.
It’s not a terrible idea, in that you can test the exact database engine binary in CI, and it’s (by definition) not a single point of failure.
I think you're oversimplifying the problem they had, and I would encourage you to dive in to the details in the article. There wasn't a problem with the database, it was with the query used to generate the configs. So if an analogous issue arose with a query against one of many ad-hoc replicated sqlite databases, you'd still have the failure.
I love sqlite for some things, but it's not The One True Database Solution.
When a failsafe system fails, it fails by failing to fail safely.
Global configuration is useful for low response times to attacks, but you need to have very good ways to know when a global config push is bad and to be able to rollback quickly.
In this case, the older proxy's "fail-closed" categorization of bot activity was obviously better than the "fail-crash", but every global change needs to be carefully validated to have good characteristics here.
Having a mapping of which services are downstream of which other service configs and versions would make detecting global incidents much easier too, by making the causative threads of changes more apparent to the investigators.
It seems they had this continous rollout for the config service, but the services consuming this were affected even by small percentage of these config providers being faulty, since they were auto updating every few minutes their configs. And it seems there is a reason for these updating so fast, presumably having to react to threat actors quickly.
It's in everyone's interest to mitigate threats as quickly as possible. But it's of even greater interest that a core global network infrastructure service provider not DOS a significant proportion of the Internet by propagating a bad configuration too quickly. The key here is to balance responsiveness against safety, and I'm not sure they struck the right balance here. I'm just glad that the impact wasn't as long and as severe as it could have been.
This isn't really "configuration" so much as it is "durable state" within the context of this system.
In my 30 years of reliability engineering, I've come to learn that this is a distinction without a difference.
People think of configuration updates (or state updates, call them what you will) as inherently safer than code updates, but history (and today!) demonstrates that they are not. Yet even experienced engineers will allow changes like these into production unattended -- even ones who wouldn't dare let a single line of code go live without being subject to the full CI/CD process.
They narrowed down the actual problem to some Rust code in the Bot Management system that enforced a hard limit on the number of configuration items by returning an error, but the caller was just blindly unwrapping it.
A dormant bug in the code is usually a condition precedent to incidents like these. Later, when a bad input is given, the bug then surfaces. The bug could have laid dormant for years or decades, if it ever surfaced at all.
The point here remains: consider every change to involve risk, and architect defensively.
They made the classic distributed systems mistake and actually did something. Never leap to thing-doing!
If they're going to yeet configs into production, they ought to at least have plenty of mitigation mechanisms, including canary deployments and fault isolation boundaries. This was my primary point at the root of this thread.
And I hope fly.io has these mechanisms as well :-)
We've written at long, tedious length about how hard this problem is.
Have a link?
Most recently, a few weeks ago (but you'll find more just a page or two into the blog):
https://fly.io/blog/corrosion/
It's great that you're working on regionalization. Yes, it is hard, but 100x harder if you don't start with cellular design in mind. And as I said in the root of the thread, this is a sign that CloudFlare needs to invest in it just like you have been.
I recoil from that last statement not because I have a rooting interest in Cloudflare but because the last several years of working at Fly.io have drilled Richard Cook's "How Complex Systems Fail"† deep into my brain, and what you said runs aground of Cook #18: Failure free operations require experience with failure.
If the exact same thing happens again at Cloudflare, they'll be fair game. But right now I feel people on this thread are doing exactly, precisely, surgically and specifically the thing Richard Cook and the Cook-ites try to get people not to do, which is to see complex system failures as predictable faults with root causes, rather than as part of the process of creating resilient systems.
† https://how.complexsystems.fail/
Suppose they did have the cellular architecture today, but every other fact was identical. They'd still have suffered the failure! But it would have been contained, and the damage would have been far less.
Fires happen every day. Smoke alarms go off, firefighters get called in, incident response is exercised, and lessons from the situation are learned (with resulting updates to the fire and building codes).
Yet even though this happens, entire cities almost never burn down anymore. And we want to keep it that way.
As Cook points out, "Safety is a characteristic of systems and not of their components."
What variant of cellular architecture are you referring to? Can you give me a link or few? I'm fascinated by it and I've led a team to break up a monolithic solution running on AWS to a cellular architecture. The results were good, but not magic. The process of learning from failures did not stop, but it did change (for the better).
No matter what architecture, processes, software, frameworks, and systems you use, or how exhaustively you plan and test for every failure mode, you cannot 100% predict every scenario and claim "cellular architecture fixes this". This includes making 100% of all failures "contained". Not realistic.
If your AWS service is properly regionalized, that’s the minimum amount of cellular architecture required. Did your service ever fail in multiple regions simultaneously?
Cellular architecture within a region is the next level and is more difficult, but is achievable if you adhere to the same principles that prohibit inter-regional coupling:
https://docs.aws.amazon.com/wellarchitected/latest/reducing-...
https://docs.aws.amazon.com/wellarchitected/latest/reducing-...
You didn't really put any thought into what I said. Thanks for the links.
It wasn't worth thinking about. I'm not going to defend myself against arguments and absolute claims I didn't make. The key word here is mitigation, not perfection.
> If your AWS service is properly regionalized, that’s the minimum amount of cellular architecture required
Amazon has had multi-region outages due to pushing bad configs, so it’s extremely difficult to believe whatever you are proposing solves that exact problem by relying on multi-regions.
Come to think of it, Cloudflare’s outage today is another good counterexample.
It has been a very, very long time since AWS had a simultaneous failure across multiple regions. Even customers impacted by the loss of Route 53 control plane functionality in last month’s us-east-1 were able to gracefully fail over to a backup region if they configured failover records in advance, had Application Recovery Controller set up, or fronted their APIs or websites with Global Accelerator.
Customers survive incidents on a daily basis by failing over across regions (even in the absence of an AWS regional failure, they can fail due to a bad deployment or other cause). The reason you don’t hear about it is because it works.
Pretty sure he's making my point (or, rather, me his) there. (I'm never going to turn down an opportunity to nerd out about Cookism).
Sounds like lack of good testing. Too many items in any input should be a boundary case you will get to eventually.
Reframe this problem: instead of bot rules being propagated, it's the enrollment of a new customer or a service at an existing customer --- something that must happen at Cloudflare several times a second. Does it still make sense to you to think about that in terms of "pushing new configuration to prod"?
Those aren't the facts before us. Also, CRUD operations relating to a specific customer or user tend not to cause the sort of widespread incidents we saw today.
They're not, they're a response to your claim that "state" and "configuration" are indifferentiable.
it's always a config push. people rollout code slowly but don't have the same mechanisms for configs. But configs are code, and this is a blind spot that causes an outsized percentage of these big outages.
An unwrap like that in production code on the critical path is very surprising to me.
I haven’t worked in Rust codebases, but I have never worked in a Go codebase where a `panic` in such a location would make it through code review.
Is this normal in Rust?
absolutely not normal, this is why in my opinion it took them so long to understand the core issue. instead of a nice error message and a backtrace saying something like "failed to parse config feature names" they thought they were under attack because the service was just crashing instead.
"Throwing us off and making us believe this might have been an attack was another apparent symptom we observed: Cloudflare’s status page went down. The status page is hosted completely off Cloudflare’s infrastructure with no dependencies on Cloudflare. While it turned out to be a coincidence, it led some of the team diagnosing the issue to believe that an attacker may be targeting both our systems as well as our status page."
Unfortunately they do not share, what caused the status page to went down as well. (Does this happen often? Otherwise a big coincidence it seems)
We don’t know. Suspect it may just have been a big uptick in load and a failure of its underlying infrastructure to scale up.
The status page is hosted on AWS Cloudfront, right? It sure looks like Cloudfront was overwhelmed by the traffic spike, which is a bit concerning. Hope we'll see a post from their side.
It looks a lot like a CloudFront error we randomly saw today from one of our engineers in South America. I suspect there was a small outage in AWS but can't prove it.
Probably non zero number of companies use cloudfront and other cdns as fallback for cloudflare or running a blended cdn so not surprising to see other cdns hit with a thundering herd when cloudflare went down
Yes, probably a bunch of automated bots decided to check the status page when they saw failures in production.
it seems like a good chance that despite thinking their status page was completely independent of cloudfront, enough of the internet is dependent on cloudfront now that they're simply wrong about the status page's independence.
i think you've got cloudflare and cloudfront mixed up.
ahah oops. yeah, it's a problem. i've got two projects ongoing that each rely on one of them, and i can never keep it straight.
Quite possibly it was due to high traffic.
IDK Atlassian Statuspage clientele, but it's possible Cloudflare is much larger than usual.
I mean, that would require a postmortem from statuspage.io right? Is that a service operated by cloudflare?
Atlasaian
Everyone is hating on unwrap, but to me the odd and more interesting part is that it took 3 hours to figure this out? Even with a DDoS red herring, shouldn’t there have been a crash log or telemetry anomaly correlated? Also, shouldn’t the next steps and resolution focus more on this aspect, since it’s a high leverage tool for identifying any outage caused by a panic rather than just preventing a recurrence of random weird edge case #9999999?
Yes this is the weird part for me. With good monitoring, the panic at unwrap should have been detected immediately. I assume they weren't looking at the right place, but still. If you use Sentry for example, a brand new panic should be pretty visible.
I have nowhere near the experience managing such complex systems, but I can empathize with this. In a high-pressure situations the most obvious things get missed. If someone is convinced System X is at fault, your mind can make leaps to justify every other degraded system is a downstream effect of that. Cause and effect can get switched.
Sometimes you have smart people in the room who dig deeper and fish it out, but you cannot always rely on that.
I have plenty of empathy, having been in plenty of similar situations. It's not a matter of "I can't BELIEVE it took that long" (although it is a bit surprising) so much as that I disagree with the key takeaways here in the HN comments section and in the blog itself, which focus strongly on fixing rare edge case issues (the bad ClickHouse query and a bad config file causing a panic via unwrap), rather than reducing MTTR for all issues by improving the debug and monitoring experience.
I'm also suspicious that
> Eliminating the ability for core dumps or other error reports to overwhelm system resources
from the blog had a lot more to do with the issue than perhaps the narrative is letting on.
Once they figured it out they didn't have a way to load in a new feature file, had to figure that out, and then restart every machine.
This took one of the three hours; it seems to have taken from 11:28 to 13:37 to recognize that the configuration file panic was the cause of the issue.
Why does cloudflare allow unwraps in their code? I would've assumed they'd have clippy lints stopping that sort of thing. Why not just match with { ok(value) => {}, Err(error) => {} } the function already has a Result type.
At the bare minimum they could've used an expect("this should never happen, if it does database schema is incorrect").
The whole point of errors as values is preventing this kind of thing.... It wouldn't have stopped the outage but it would've made it easy to diagnose.
If anyone at cloudflare is here please let me in that codebase :)
And the error magically disappears when the function returns it?
It doesn’t disappear, it forces you to handle it.
Not a cloudflare employee but I do write a lot of Rust. The amount of things that can go wrong with any code that needs to make a network call is staggeringly high. unwrap() is normal during development phase but there are a number of times I leave an expect() for production because sometimes there's no way to move forward.
Yeah it seems likely that even if there wasn't an unwrap, there would have been some error handling that wouldn't have panicked the process, but would have still left it inoperable if every request was instead going through an error path.
I'm in a similar boat, at the very leas an expect can give hits to what happened. However this can also be problematic if your a library developer. Sometimes rust is expected to never panic especially in situations like WASM. This is a major problem for companies like Amazon Prime Video since they run in a WASM context for their TV APP. Any panic crashes everything. Personally I usually just either create a custom error type (preferred) or erase it away with Dyn Box Error (no other option). Random unwraps and expects haunt my dreams.
Long time ago Google had a very similar incident where ddos protection system ingested a bad config and took everything down. Except it was auto resolved in like four minutes by an automatic rollback system before oncall was even able to do anything. Perhaps Cloudflare should invest in a system like that
The most surprising thing to me here is that it took 3 hours to root cause, and points to a glaring hole in the platform observability. Even taking into account the fact that the service was failing intermittently at first, it still took 1.5 hours after it started failing consistently to root cause. But the service was crashing on startup. If a core service is throwing a panic at startup like that, it should be raising alerts or at least easily findable via log aggregation. It seems like maybe there was some significant time lost in assuming it was an attack, but it also seems strange to me that nobody was asking "what just changed?", which is usually the first question I ask during an incident.
> thread fl2_worker_thread panicked: called Result::unwrap() on an Err value
I don't use Rust, but a lot of Rust people say if it compiles it runs.
Well Rust won't save you from the usual programming mistake. Not blaming anyone at cloudflare here. I love Cloudflare and the awesome tools they put out.
end of day - let's pick languages | tech because of what we love to do. if you love Rust - pick it all day. I actually wanna try it for industrial robot stuff or small controllers etc.
there's no bad language - just occassional hiccups from us users who use those tools.
You misunderstand what Rust’s guarantees are. Rust has never promised to solve or protect programmers from logical or poor programming. In fact, no such language can do that, not even Haskell.
Unwrapping is a very powerful and important assertion to make in Rust whereby the programmer explicitly states that the value within will not be an error, otherwise panic. This is a contract between the author and the runtime. As you mentioned, this is a human failure, not a language failure.
Pause for a moment and think about what a C++ implementation of a globally distributed network ingress proxy service would look like - and how many memory vulnerabilities there would be… I shudder at the thought… (n.b. nginx)
This is the classic example of when something fails, the failure cause over indexes on - while under indexing on the quadrillions of memory accesses that went off without a single hitch thanks to the borrow checker.
I postulate that whatever the cost in millions or hundreds of millions of dollars by this Cloudflare outage, it has paid for more than by the savings of safe memory access.
See: https://en.wikipedia.org/wiki/Survivorship_bias
> You misunderstand what Rust’s guarantees are.
Well, no, most Rust programmers misunderstand what the guarantees are because they keep parroting this quote. Obviously the language does not protect you from logic errors, so saying "if it compiles, it works" is disingenuous, when really what they mean is "if it compiles, it's probably free of memory errors".
No, the "if it compiles, it works" is genuinely about the program being correct rather than just free of memory errors, but it's more of a hyperbolic statement than a statement of fact.
It's a common thing I've experienced and seen a lot of others say that the stricter the language is in what it accepts the more likely it is to be correct by the time you get it to run. It's not just a Rust thing (although I think Rust is _stricter_ and therefore this does hold true more of the time), it's something I've also experienced with C++ and Haskell.
So no, it's not a guarantee, but that quote was never about Rust's guarantees.
> Pause for a moment and think about what a C++ implementation of a globally distributed network ingress proxy service would look like - and how many memory vulnerabilities there would be… I shudder at the thought
I mean thats an unfalsifiable statement, not really fair. C is used to successfully launch spaceships.
Whereas we have a real Rust bug that crashed a good portion of the internet for a significant amount of time. If this was a C++ service everyone would be blaming the language, but somehow Rust evangelicals are quick to blame it on "unidiomatic Rust code".
A language that lets this easily happen is a poorly designed language. Saying you need to ban a commonly used method in all production code is broken.
Only formal proof languages are immune to such properties. Therefore all languages are poorly designed by your metric.
Consider that the set of possible failures enabled by language design should be as small as possible.
Rust's set is small enough while also being productive. Until another breakthrough in language design as impactful as the borrow checker is invented, I don't imagine more programmers will be able to write such a large amount of safe code.
I would say the impact of the borrow checker is exaggerated.
> Rust won't save you from the usual programming mistake.
Disagree. Rust is at least giving you an "are you sure?" moment here. Calling unwrap() should be a red flag, something that a code reviewer asks you to explain; you can have a linter forbid it entirely if you like.
No language will prevent you from writing broken code if you're determined to do so, and no language is impossible to write correct code in if you make a superhuman effort. But most of life happens in the middle, and tools like Rust make a huge difference to how often a small mistake snowballs into a big one.
> Disagree. Rust is at least giving you an "are you sure?" moment here. Calling unwrap() should be a red flag, something that a code reviewer asks you to explain; you can have a linter forbid it entirely if you like.
No one treats it like that and nearly every Rust project is filled with unwraps all over the place even in production system like Cloudflare's.
Well let me avoid those that don’t understand it. It’s literally Rust 101.
I've worked on commercial codebases that did better, shrug.
> you can have a linter forbid it entirely if you like.
It would be better if that would be the other way round "linter forbids it unless you ask it not to". Never wrong to allow users to shoot themself in the foot, but it should be explicit.
Yep, unwrap() and unsafe are escape hatches that need very good justifications. It's fine for casual scripts where you don't care if it crashes. For serious production software they should be either banned, or require immense scrutiny.
What people are saying is that idiomatic prod rust doesn't use unwrap/expect (both of which panic on the "exceptional" arm of the value) --- instead you "match" on the value and kick the can up a layer on the call chain.
What happens to it up the callstack? Say they propagated it up the stack with `?`. It has to get handled somewhere. If you don't introduce any logic to handle the duplicate databases, what else are you going to do when the types don't match up besides `unwrap`ing, or maybe emitting a slightly better error message? You could maybe ignore that module's error for that request, but if it was a service more critical than bot mitigation you'd still have the same symptom of getting 500'd.
> What happens to it up the callstack?
as they say in the post, these files get generated every 5 minutes and rolled out across their fleet.
so in this case, the thing farther up the callstack is a "watch for updated files and ingest them" component.
that component, when it receives the error, can simply continue using the existing file it loaded 5 minutes earlier.
and then it can increment a Prometheus metric (or similar) representing "count of errors from attempting to load the definition file". that metric should be zero in normal conditions, so it's easy to write an alert rule to notify the appropriate team that the definitions are broken in some way.
that's not a complete solution - in particular it doesn't necessarily solve the problem of needing to scale up the fleet, because freshly-started instances won't have a "previous good" definition file loaded. but it does allow for the existing instances to fail gracefully into a degraded state.
in my experience, on a large enough system, "this could never happen, so if it does it's fine to just crash" is almost always better served by a metric for "count of how many times a thing that could never happen has happened" and a corresponding "that should happen zero times" alert rule.
Given that the bug was elsewhere in the system (the config file parser spuriously failed), it’s hard to justify much of what you suggested.
Panics should be logged, and probably grouped by stack trace for things like prometheus (outside of process). That handles all sorts of panic scenarios, including kernel bugs and hardware errors, which are common at cloudflare scale.
Similarly, mitigating by having rapid restart with backoff outside the process covers far more failure scenarios with far less complexity.
One important scenario your approach misses is “the watch config file endpoint fell over”, which probably would have happened in this outage if 100% of servers went back to watching all of a sudden.
Sure, you could add an error handler for that too, and for prometheus is being slow, and an infinite other things. Or, you could just move process management and reporting out of process.
The way I’ve seen this on a few older systems was that they always keep the previous configuration around so it can switch back. The logic is something like this:
1. At startup, load the last known good config.
2. When signaled, load the new config.
3. When that passes validation, update the last-known-good pointer to the new version.
That way something like this makes the crash recoverable on the theory that stale config is better than the service staying down. One variant also recorded the last tried config version so it wouldn’t even attempt to parse the latest one until it was changed again.
For Cloudflare, it’d be tempting to have step #3 be after 5 minutes or so to catch stuff which crashes soon but not instantly.
Presumably you kick up the error to a level that says “if parsing new config fails, keep the old config”
The config file subsystem was where the bug lived, not the code with the unwrap, so this sort of change is a special case of “make the unwrap never fail and then fix the API so it is not needed”.
Yeah, see, that's what I mean.
> Well Rust won't save you from the usual programming mistake
This is not a Rust problem. Someone consciously chose to NOT handle an error, possibly thinking "this will never happen". Then someone else conconciouly reviewed (I hope so) a PR with an unwrap() and let it slide.
And people doing testing failed to ignore their excuse of this never happening and still testing it. With this kind of systems you need the separate group that just ignores any "this will never happen" and still checks what happens if it does.
Now it might be that it was tested, but then ignored or deprioritised by management...
> I don't use Rust, but a lot of Rust people say if it compiles it runs.
Do you grok what the issue was with the unwrap, though...?
Idiomatic Rust code does not use that. The fact that it's allowed in a codebase says more about the engineering practices of that particular project/module/whatever. Whoever put the `unwrap` call there had to contend with the notion that it could panic and they still chose to do it.
It's a programmer error, but Rust at least forces you to recognize "okay, I'm going to be an idiot here". There is real value in that.
While I agree that Rust got it right by being more explicit, a lot of bugs in C/C++ can also easily avoided with good engineering practices. The Rust argument that it is mainly the fault of the programming language with C/C++ was always a huge and unfair exaggeration. Now with this entirely predictable ".unwrap" desaster (in general, not necessarily this exact scenarious), the "no true Rustacean would have put unwrap in production" fallacy is sad and funny at the same time.
> the "no true Rustacean would have put unwrap in production"
The "no unwrap" rule is common in most production codebases. Chill.
Unwrap is controversial. The problem is that if you remove it, it makes the bar even higher for newcomers to Rust. One solution is to make it unsafe (along with panic).
other people might say - why use unsafe rust - but we don't know the conditions of what the original code shipped under. why the pr was approved.
could have been tight deadline, managerial pressure or just the occasional slip up.
We shouldn't be having critical internet-wide outages on a monthly basis. Something is systematically wrong with the way we're architecting our systems.
Cloudflare, Azure, and other single points of failure are solving issues inherent to webhosting, and those problems have become incredibly hard due to the massive scale of bad actors and the massive complexity of managing hardware and software.
What would you propose to fix it? The fixed cost of being DDoS-proof is in the hundreds of millions of dollars.
Costs to architect systems that serve millions of request daily have gone down. Not up.
Hell, I would be very curious to know the costs to keep HackerNews running. They probably serve more users than my current client.
People want to chase the next big thing to write it on their CV, not architect simple systems that scale. (Do they even need to scale?)
> Costs to architect systems that serve millions of request daily have gone down. Not up.
I never said serving millions of requests is more expensive. Protecting your servers is more expensive.
> Hell, I would be very curious to know the costs to keep HackerNews running. They probably serve more users than my current client.
HN uses Cloudflare. You're making my point for me. If you included the fixed costs that Cloudflare's CDN/proxy is giving to HN incredibly cheaply, then running HN at the edge with good performance (and protecting it from botnets) would costs hundreds of millions of dollars.
> People want to chase the next big thing to write it on their CV, not architect simple systems that scale. (Do they even need to scale?)
Again, attacking your own straw men here.
Writing high-throughput web applications is easier than ever. Hosting them on the open web is harder than ever.
I don't understand what's the business of cloudflare.
They just sell proxies, to whoever.
Why are they the only company doing ddos protection?
I just don't get it.
Momentum, I guess, pretty much like you wouldn’t get fired for using AWS or IBM (if that’s still the case now).
While I heavily frown upon using `unwrap` and `expect` in Rust code and make sure to have Clippy tell me about every single usage of them, I also understand that without them Rust might have been seen as an academic curiosity language.
They are escape hatches. Without those your language would never take off.
But here's the thing. Escape hatches are like emergency exits. They are not to be used by your team to go to lunch in a nearby restaurant.
---
Cloudflare should likely invest in better linting and CI/CD alerts. Not to mention isolated testing i.e. deploy this change only to a small subset and monitor, and only then do a wider deployment.
Hindsight is 20/20 and we can all be smartasses after the fact of course. But I am really surprised because lately I am only using Rust for hobby projects and even I know I should not use `unwrap` and `expect` beyond the first iteration phases.
---
I have advocated for this before but IMO Rust at this point will benefit greatly from disallowing those unsafe APIs by default in release mode. Though I understand why they don't want to do it -- likely millions of CI/CD pipelines will break overnight. But in the interim, maybe a rustc flag we can put in our `Cargo.toml` that enables such a stricter mode? Or have that flag just remove all the panicky API _at compile time_ though I believe this might be a Gargantuan effort and is likely never happening (sadly).
In any case, I would expect many other failures from Cloudflare but not _this_ one in particular.
This is not a reasonable take to me. unwrap/expect are the idiomatic way to express code paths returning Option/Result as unreachable.
Bubbling up the error or None does not make the program correct. Panicking may be the only reasonable thing to do.
If panicking is guaranteed because of some input mistake to the system your failure is in testing.
Say you panic, now you need to be an external system that catches this panic and report back; and do something meanwhile to recover your system.
If you think about it, it’s not really different from handling the bubbled up error inside of Rust. You don’t (?) your results and your errors go away, they just move up the chain.
I agree the failure is in testing but what you can and should do is raise in alert in your APM system before the runtime panic, in the code path that is deemed impossible to hit.
I am not trashing on them, I've made such mistakes in the past, but I do expect more from them is all.
And you will not believe how many alerts I got for the "impossible" errors.
I do agree there was not too much that could have been done, yes. But they should have invested in more visibility and be more thorough. I mean, hobbyist Rust devs seem to do that better.
It was just a bit disappointing for me. As mentioned above, I'd understand and sympathise with many other mistakes but this one stung a bit.
There's certainly a discipline involved here, but it's usually something like guaranteeing all threads are unwind safe (via AssertUnwindSafe) and logging stack traces when your process keeps dying/can't be started after a fixed number of retries. Which would lead you to the culprit immediately.
I'm just pushing back a bit on the idea that unwrap() is unsafe - it's not, and I wouldn't even call it a foot gun. The code did what it was written to do, when it saw the input was garbage it crashed because it couldn't make sense of what to do next. That's a desirable property in reliable systems (of course monitoring that and testing it is what makes it reliable/fixable in the first place).
We don't disagree, my main point was a bit broader and admittedly hijacked the original topic a bit, namely: `unwrap` and `expect` make many Rust devs too comfortable and these two are very tempting mistresses.
Using those should be done in an extremely disciplined manner. I agree that there are many legitimate uses but in the production Rust code I've seen this has rarely been the case. People just want to move on and then forget to circle back and add proper error handling. But yes, in this case that's not quite true. Still, my point that an APM alert should have been raised on the "impossible" code path before panicking, stands.
Oh for sure. I even think there deserve to be lints like "no code path reachable from main() is unwind-unsafe" which is a heavy hammer for many applications (like one-off CLI utils) but absolutely necessary for something like a long-lived daemon or server that's responsible for critical infrastructure.
Sure this has been said but the issue here is not code is the ability to canary and rollback quickly from any arbitrary (config) change.
Lots of people here are (perhaps rightfully) pointing to the unwrap() call being an issue. That might be true, but to me the fact that a reasonably "clean" panic at a defined line of code was not quickly picked up in any error monitoring system sounds just as important to investigate.
Assuming something similar to Sentry would be in use, it should clearly pick up the many process crashes that start occurring right as the downtime starts. And the well defined clean crashes should in theory also stand out against all the random errors that start occuring all over the system as it begins to go down, precisely because it's always failing at the exact same point.
Exactly! You could have `rand() > 0.5 && panic!()` in the code of your bot module, and that should not put the internet on fire.
The issue here is about the system as a whole not any line of code.
Wow. 26M/s 5xx error HTTP status codes over a span of roughly two hours. That's roughly 187 billion HTTP errors that interrupted people (and systems)!
I integrated Turnstile with a fail-open strategy that proved itself today. Basically, if the Turnstile JS fails to load in the browser (or in a few specific frontend error conditions), we allow the user to submit the web form with a dummy challenge token. On the backend, we process the dummy token like normal, and if there is an error or timeout checking Turnstile's siteverify endpoint, we fail open.
Of course, some users were still blocked, because the Turnstile JS failed to load in their browser but the subsequent siteverify check succeeded on the backend. But overall the fail-open implementation lessened impact to our customers nonetheless.
Fail-open with Turnstile works for us because we have other bot mitigations that are sufficient to fall back on in the event of a Cloudflare outage.
So to bypass captcha all a user has to do is block the script from loading? I can see that working but only for attacks that aren’t targeted?
Only if they are able to block the siteverify check performed by our backend server. That's not the kind of attack we are trying to mitigate with Turnstile.
Classic combination of errors:
Having the feature table pivoted (with 200 feature1, feature2, etc columns) meant they had to do meta queries to system.columns to get all the feature columns which made the query sensitive to permissioning changes (especially duplicate databases).
A Crowdstrike style config update that affects all nodes but obviously isn't tested in any QA or staged rollout strategy beforehand (the application panicking straight away with this new file basically proves this).
Finally an error with bot management config files should probably disable bot management vs crash the core proxy.
I'm interested here why they even decided to name Clickhouse as this error could have been caused by any other database. I can see though the replicas updating causing flip / flopping of results would have been really frustrating for incident responders.
Right but also this is a pretty common pattern in distributed systems that publish from databases (really any large central source of truth); it might be like the problem in systems like this. When you're lucky the corner cases are obvious; in the big one we experienced last year, a new row in our database tripped an if-let/mutex deadlock, which our system dutifully (and very quickly) propagated across our entire network.
The solution to that problem wasn't better testing of database permutations or a better staging environment (though in time we did do those things). It was (1) a watchdog system in our proxies to catch arbitrary deadlocks (which caught other stuff later), (2) segmenting our global broadcast domain for changes into regional broadcast domains so prod rollouts are implicitly staged, and (3) a process for operators to quickly restore that system to a known good state in the early stages of an outage.
(Cloudflare's responses will be different than ours, really I'm just sticking up for the idea that the changes you need don't follow obviously from the immediate facts of an outage.)
But Rust was supposed to cure cancer and solve world hunger. Is this the end of the hello world but in Rust saga?
Cloudflare Access is still experiencing weird issues for us (it’s asking users to SSO login to our public website even though our zone rules - set on a completely different zone - haven’t changed).
I don’t think the infrastructure has been as fully recovered as they think yet…
People really like to hate on Rust for some reason. This wasn’t a Rust problem, no language would have saved them from this kind of issue. In fact, the compiler would have warned that this was a possible issue.
I get it, don’t pick languages just because they are trendy, but if any company’s use case is a perfect fit for Rust it’s cloudflare.
Yeah even if you handled this situation without unwrap() if you just went down an error path that didn't panic, the service would likely still be inoperable if every single request went down the error path.
Okay, but if you returned a wrapped error it’d at least be easier to debug.
The reason why people are criticizing is because Rust evangelicals say stuff like "if it compiles it works" or talk about how Rust's type system is so much better than other languages that it catches logic errors like this. You won't see Go or Java developers making such strong claims about their preferred languages.
> [...] it catches logic errors like this
but Rust's type system did catch this error - and then author decided it's fine to panic if this error happens
> You won't see Go or Java developers making such strong claims about their preferred languages.
yess no Java developer ever said that OOP will solve world hunger
> but Rust's type system did catch this error - and then author decided it's fine to panic if this error happens
The issue is that it wasn't fine to panic, thus Rust did not catch this error.
Having a system which automatically deploys configuration files across a million servers every 5 minutes without testing it seems stupid to me.
The real take away is that so much functionality depends on a few players. This is a fundamental flaw in design that is getting worse by the year as the winner takes all winners win. Not saying they didn’t earn their wins. But the fact remains. The system is not robust. Then again, so what. It went down for a while. Maybe we shouldn’t depend on the internet being “up” all the time.
"Customers on our old proxy engine, known as FL, did not see errors, but bot scores were not generated correctly, resulting in all traffic receiving a bot score of zero."
This simply means, the exception handling quality of your new FL2 is non-existent and is not at par / code logic wise similar to FL.
I hope it was not because of AI driven efficiency gains.
In most domains, silently returning 0 in a case where your logic didn't actually calculate the thing you were trying to calculate is far worse than giving a clear error.
A lot of outages off late seem to be related to automated config management.
Companies seem to place a lot of trust is configs being pushed automatically without human review into running systems. Considering how important these configs are, shouldn't they perhaps first be deployed to a staging/isolated network for a monitoring window before pushing to production systems?
Not trying to pontificate here, these systems are more complicated than anything I have maintained. Just trying to think of best practices perhaps everyone can adopt.
Why call .unwrap() in a function which returns Result<_,_>?
For something so critical, why aren't you using lints to identify and ideally deny panic inducing code. This is one of the biggest strengths of using Rust in the first place for this problem domain.
Probably because this case was something more akin to an assert than an error check.
Fly writes a lot of Rust, do you allow `unwrap()` in your production environment? At Modal we only allow `expect("...")` and the message should follow the recommended message style[1].
I'm pretty surprised that Cloudflare let an unwrap into prod that caused their worst outage in 6 years.
1. https://doc.rust-lang.org/std/option/enum.Option.html#recomm...
After The Great If-Let Outage Of 2024, we audited all our code for that if-let/rwlock problem, changed a bunch of code, and immediately added a watchdog for deadlocks. The audit had ~no payoff; the watchdog very definitely did.
I don't know enough about Cloudflare's situation to confidently recommend anything (and I certainly don't know enough to dunk on them, unlike the many Rust experts of this thread) but if I was in their shoes, I'd be a lot less interested in eradicating `unwrap` everywhere and more in making sure than an errant `unwrap` wouldn't produce stable failure modes.
But like, the `unwrap` thing is all programmers here have to latch on to, and there's a psychological self-soothing instinct we all have to seize onto some root cause with a clear fix (or, better yet for dopaminergia, an opportunity to dunk).
A thing I really feel in threads like this is that I'd instinctively have avoided including the detail about an `unwrap` call --- I'd have worded that part more ambiguously --- knowing (because I have a pathological affinity for this community) that this is exactly how HN would react. Maybe ironically, Prince's writing is a little better for not having dodged that bullet.
Sounds like if nothing else, additional attention around (their?) use of unwrap() is still warranted from where you're sitting then though, no? I don't think there's anything wrong with flagging that.
It's one thing to not want to be the one to armchair it, but that doesn't mean that one has to suppress their normal and obvious reactions. You're allowed to think things even if they're kitsch, you too are human, and what's kitsch depends and changes. Applies to everyone else here by extension too.
Fair. I agree that saying "it's the unwrap" and calling it a day is wrong. Recently actually we've done an exercise on our Worker which is "assume the worst kind of panic happens. make the Worker be ok with it".
But I do feel strongly that the expect pattern is a highly useful control and that naked unwraps almost always indicate a failure to reason about the reliability of a change. An unwrap in their core proxy system indicates a problem in their change management process (review, linting, whatever).
Rust has debug asserts for that. Using expect with a comment about why the condition should not/can't ever happen is idiomatic for cases where you never expect an Err.
This reads to me more like the error type returned by append with names is not (ErrorFlags, i32) and wasn't trivially convertible into that type so someone left an unwrap in place on an "I'll fix it later" basis, but who knows.
Oh absolutely, that's how it would have been treated.
Surely a unwrap_or_default() would have been a much better fit--if fetching features fails, continue processing with an empty set of rules vs stop world.
In that case, it should probably be expect rather than unwrap, to document why the assertion should never fail.
You are saying this would not have happened in a C release build where asserts define to nothing?
Wonder why these old grey beards chose to go with that.
I am one of those old grey beards (or at least, I got started shipping C code in the 1990s), and I'd leave asserts in prod serverside code given the choice; better that than a totally unpredictable error path.
> You are saying this would not have happened in a C release build where asserts define to nothing?
Afaik, Go and Java are the only languages that make you pause and explicitly deal with these exceptions.
And rust, but they chose to panic on the error condition. Wild.
> And rust, but they chose to panic on the error condition. Wild.
unwrap() implicitly panic-ed, right?
I don't think "implicitly panicked" is an accurate description since unwrap()'s entire reason for existing is to panic if you unwrap an error condition. If you use unwrap(), you're explicitly opting into the panicking behavior.
I suppose another way to think about it is that Result<T, E> is somewhat analogous to Java's checked exceptions - you can't get the T out unless you say what to do in the case of the E/checked exception. unwrap() in this context is equivalent to wrapping the checked exception in a RuntimeException and throwing that.
Yes, can't have .unwrap() in production code (it's ok in tests)
Like goto, unwrap is just a tool that has its use cases. No need to make a boogeyman out of it.
Yes it's meant to be used in test code. If you're sure it can't fail do then use .expect() that way it shows you made a choice and it wasn't just a dev oversight.
panicans should be using .expect() in production
To be fair, if you’re not “this tall” you really shouldn’t consider using goto in a c program. Most people aren’t that tall.
Nonsense. Linux kernel for one example, uses goto everywhere for error handling.
unwrap itself isn't the problem...
Why is there a 200 limit on appending names?
Everything has a limit. You can define it, or be surprised when you find out what it is.
Limits in systems like these are generally good. They mention the reasoning around it explicitly. It just seems like the handling of that limit is what failed and was missed in review.
The outage sucked for everyone. The root cause also feels like something they could have caught much earlier in a canary rollout from my reading of this.
All that said, to have an outage reported turned around practically the same day, that is this detailed, is quite impressive. Here's to hoping they make their changes from this learning, and we don't see this exact failure mode again.
how would you build redundancy around cloudflare failing?
i think this is happening way too frequently
meanwhile VPS, dedicated servers hum along without any issues
i dont want to use kubernetes but if we have to build mission critical systems doesn't seem like building on cloudflare is going to cut it
I'm feeling generous tonight, I'm willing to consider 0.99986 to round to 99.99%
May I just say that Matthew Prince is the CEO of Cloudflare and a lawyer by training (and a very nice guy overall). The quality of this postmortem is great but the fact that it is from him makes one respect the company even more.
https://x.com/guanlandai/status/1990967570011468071
Configuration can be validated. https://cuelang.org
I don't get why that SQL query was even used in the first place. It seems it fetches feature names at runtime instead of using a static hardcoded schema. Considering this decides the schema of a global config, I don't think the dynamicity is a good idea.
kudos to getting this blog post out so fast, it’s well written and is appreciated.
i’m a little confused on how this was initially confused for an attack though?
is there no internal visibility into where 5xx’s are being thrown? i’m surprised there isn’t some kind of "this request terminated at the <bot checking logic>" error mapping that could have initially pointed you guys towards that over an attack.
also a bit taken aback that .unwrap()’s are ever allowed within such an important context.
would appreciate some insight!
Question: customer having issues also couldn't switch their dns to bypass the service, why is the control plane updated along the data plane here it seem a lot of use could save business continuity if they could change their dns entry temporarily
Hack Free fire 8000 5487 3565 644664 464664644
464646449
Unbelievable. I guess it's time to grep for every .unwrap in our code.
theres a clippy lint for that
Time to rewrite with golang, explicit error handling ;-)
Hold up ,- when I used a C or similar language for accessing a database and wanted to clamp down on memory usage to deterministically control how much I want to allocated, I would explicitly limit the number of rows in the query.
There never was an unbound "select all rows from some table" without a "fetch first N rows only" or "limit N"
If you knew that this design is rigid, why not leverage the query to actually do it ?
What am I missing ?
Because nothing forced them to and they didn't think of it. Maybe the people writing the code that did the query knew that the tables they were working with never had more than 60 rows and figured "that's small" so they didn't bother with a limit. Maybe the people who wrote the file size limit thought "60 rows isn't that much data" and made a very small file size limit and didn't coordinate with the first people.
Anyway regardless of which language you use to construct a SQL query, you're not obligated to put in a max rows
I imagine there's numerous ways to protect against it and protection should've been added by whoever decided on this optimization. In data layer, create some kind of view which never returns more than 200 rows from base table(s). In code, use some kind of iterator. I'm not a Rust guy, just a C defensive practices type of dude, but maybe they just missed a biggie during a code review.
Reason for the failure: switched to Chad IDE to ship new features.
Cloudflare’s write-up is clear and to the point. A small change spread wider than expected, and they explained where the process failed. It’s a good reminder that reliability depends on strong workflows as much as infrastructure.
cloudflare:
> Throwing us off and making us believe this might have been an attack was another apparent symptom we observed: Cloudflare’s status page went down. The status page is hosted completely off Cloudflare’s infrastructure with no dependencies on Cloudflare.
also cloudflare:
> The Cloudflare Dashboard was also impacted due to both Workers KV being used internally and Cloudflare Turnstile being deployed as part of our login flow.
I believe you're mistakenly equating Cloudflare's status page with the Cloudflare Dashboard? They're not the same thing.
Cloudflare's status page: https://www.cloudflarestatus.com/
Cloudflare Dashboard: https://dash.cloudflare.com/
thanks for clarifying! i guess then they never explained why the status page went down, even though it's supposed to be running on independent infrastructure.
Yes, that was missing (along with the London WARP thing). Other comments mentioned that their status page is an Atlassian Statuspage solution, hosted on AWS CloudFront.
Unclear to me if it's an Atlassian-managed deployment they have, or if it's self-managed, I'm not familiar with Statuspage and their website isn't helping. Though if it's managed, I'm not sure how they can know for sure there's no interdependence. (Though I guess we could technically keep that rabbit hole going indefinitely.)
Great post-mortem. Very clear. Surprised that num(panicking threads) didn't show up somewhere in telemetry.
Would be nice if their Turnstile could be turned off on their login page when something like this happens, so we can attempt to route traffic away from Cloudflare during the outage. Or at least have a simple app where this can be modified from.
Honestly... everyone shit themselves that internet doesn't work, but next week this outage will be forgotten by 99% of population. I was doing something on my PC when I saw clear information that Cloudflare is down, so I decided to just go take a nap, then read a book, then go for a walk. Once I was done, the internet was working again. Panic was not necessary on my side.
What I'm trying to say is that things would be much better if everyone took a chill pill and accepted the possibility that in rare instances, the internet doesn't work and that's fine. You don't need to keep scrolling TikTok 24/7.
> but my use case is especially important
Take a chill pill. Probably it isn't.
Wondering why they didn’t disable the bot management temporarily to recover. Websites could have survived temporarily without it compared to the outage itself.
ClickHouse db was mentioned. Does this incident raise any critiques about it?
I would have been a bit cheeky and opened with 'It wasn't DNS.'
Was it DNS this time?
unwraps are so very easy to use and they have bit me so many times because you can nearly never run into a problem and suddenly crashes from an unwrap that almost always was fine
If you deploy a change to your system, and things start to go wrong that same day, the prime suspect (no matter how unlikely it might seem) should be the change you made.
My first question when faced with an unknown error is "What was the last change and when was it promoted?"
This is an area where they are allowed to think yet another record setting DDoS attack first and bad config second.
F
Hack
Free fire
>Currently that limit is set to 200, well above our current use of ~60 features. Again, the limit exists because for performance reasons we preallocate memory for the features.
So they basically hardcoded something, didn't bother to cover the overflow case with unit tests, didn't have basic error catching that would fallback and send logs/alerts to their internal monitoring system and this is why half of the internet went down?
This is an excellent lesson learned: Harden loading of internally generated config files as though they were untrusted content.
Gonna use that one at $WORK.
“…and the fluctuation stabilized in the failing state.”
Sounds like the ops team had one hell of a day.
Wow. What a post mortem. Rather than Monday morning quarterbacking how many ways this could have been prevented, I'd love to hear people sound-off on things that unexpectedly broke. I, for one, did not realize logging in to porkbun to edit DNS settings would become impossible with a cloudflare meltdown
That's unfortunate. I'll need to investigate whether Porkbun plans on decoupling its auth from being reliant on CloudFlare, otherwise I will need to migrate a few domains off of that registrar.
Ironically just now I got a Cloudflare "Error code 524" page because blog.cloudflare.com was down
> I worry this is the big botnet flexing.
Even worse - the small botnet that controls everything.
While it's certainly worthwhile to discuss the Technical and Procedural elements that contributed to this Service Outage, the far more important (and mutually-exclusive aspect) to discuss should be:
Why have we built / permitted the building of / Subscribed to such a Failure-intolerant "Network"?
Who's "we"? This is not a trick question, what specific people do you think acted wrongly here? I don't use Cloudflare personally. I don't run any of the sites that do use it. The people who did make the decision to put thier websites behind Cloudflare could stop, and maybe some will, but presumably they're paying for it because they think, perhaps accurately, that they get value out of it. Should some power compel them not to use Cloudflare?
Cloudflare rewrites Rust services to <next-cool-language> /joke
...
(I'd pick Haskell, cause I'm having fun with it recently :P)
Matt, Looking forward in regaining Elon's and his team trust to use CF again.
I wish Elon would regain my trust!
This post was written by chatgpt??
https://blog.cloudflare.com/18-november-2025-outage/#:~:text...
Here's a random post from their blog by the same author from 2017 with an em dash:
> As we wrote before, we believe Blackbird Tech's dangerous new model of patent trolling — where they buy patents and then act their own attorneys in cases — may be a violation of the rules of professional ethics.
https://blog.cloudflare.com/patent-troll-battle-update-doubl...
ChatGPT didn't invent the em dash, some people were always using it. But yeah, it's often one of the signs of AI.
deny (clippy:: unwrap_used)
tl;dr A permissions change in a ClickHouse database caused a query to return duplicate rows for a “feature file” used by Cloudflares Bot Management system, which doubled the file size. That oversized file was propagated to their core proxy machines, triggered an unhandled error in the proxy’s bot-module (it exceeded its pre-allocated limit), and as a result the network started returning 5xx errors. The issue wasn’t a cyber-attack — it was a configuration/automation failure.
Excellent write up. Cybersecurity professionals read the story and learn. It’s textbook lesson in post-mortem incident analysis - a mvp for what is expected from us all in a similar situation.
Reputationally this is extremely embarrassing for Cloudflare, but imo they seem to get their feet back on the ground. I was surprised to see not just one, but two apologies to the internet. This just cements how professional and dedicated the Cloudflare team is to ensure stable resilient internet and how embarrassed they must have been.
A reputational hit for sure, but outcome is lessons learned and hopefully stronger resilience.
I thought it was an internal mess-up. I thought an employee screwed a file up. Old methods are sometimes better than new. AI fails us again!
It is staggering to see that even large companies like CF have zero monitoring, so they would know what happened in t=0.
So an unhandled error condition after an configuration update similar to Crowdstrike - if they had just used a programming language where this can't happen due to the superior type system such as Rust. Oh wait.
> Instead, it was triggered by a change to one of our database systems' permissions which caused the database to output multiple entries into a “feature file” used by our Bot Management system.
And here is the query they used ** (OK, so it's not exactly):
someone added a new row to permissions and the JOIN started returning two dupe feature rows for each distinct feature.** "here is the query" is used for dramatic effect. I have no knowledge of what kind of database they are even using much less queries (but i do have an idea).
more edits: OK apparently it's described later in the post as a query against clickhouse's table metadata table, and because users were granted access to an additional database that was actually the backing store to the one they normally worked with, some row level security type of thing doubled up the rows. Not sure why querying system.columns is part of a production level query though, seems overly dynamic.
I believe they mentioned ClickHouse
Wow, crazy disproportional drop in the stock price… good buying opportunity for $NET.
Agree.
Cloudflare is very cheap at these prices.
I can never get used to the error happening at call site rather than within the function where the early return of Err happened. It is not "much cleaner", you have no idea which line and file caused it at call site. By default Returning should have a way of setting a marker which can then be used to map back to the line() and file(). 10+ years and still no ergonomics.
> a change to one of our database systems' permissions which caused the database to output multiple entries into a “feature file” used by our Bot Management system ... to keep [that] system up to date with ever changing threats
> The software had a limit on the size of the feature file that was below its doubled size. That caused the software to fail
A configuration error can cause internet-scale outages. What an era we live in
Edit: also, after finishing my reading, I have to express some surprise that this type of error wasn't caught in a staging environment. If the entire error is that "during migration of ClickHouse nodes, the migration -> query -> configuration file pipeline caused configuration files to become illegally large", it seems intuitive to me that doing this same migration in staging would have identified this exact error, no?
I'm not big on distributed systems by any means, so maybe I'm overly naive, but frankly posting a faulty Rust code snippet that was unwrapping an error value without checking for the error didn't inspire confidence for me!
It would have been caught only in stage if there was similar amount of data in the database. If stage has 2x less data it would have never occurred there. Not super clear how easy it would have been to keep stage database exactly as production database in terms of quantity and similarity of data etc.
I think it's quite rare for any company to have exact similar scale and size of storage in stage as in prod.
> I think it's quite rare for any company to have exact similar scale and size of storage in stage as in prod.
We’re like a millionth the size of cloudflare and we have automated tests for all (sort of) queries to see what would happen with 20x more data.
Mostly to catch performance regressions, but it would work to catch these issues too.
I guess that doesn’t say anything about how rare it is, because this is also the first company at which I get the time to go to such lengths.
But now consider how much extra data Cloudflare at its size would have to have just for staging, doubling or more their costs to have stage exactly as production. They would have to simulate similar amount of requests on top of themselves constantly since presumably they have 100s or 1000s of deployments per day.
In this case it seems the database table in question seemed modest in size (the features for ML) so naively thinking they could have kept stage features always in sync with prod at the very least, but could be they didn't consider that 55 rows vs 60 rows or similar could be a breaking point given a certain specific bug.
It is much easier to test with 20x data if you don't have the amount of data cloudflare probably handles.
That just means it takes longer to test. It may not be possible to do it in a reasonable timeframe with the volumes involved, but if you already have 100k servers running to serve 25M requests per second, maybe briefly booting up another 100k isn’t going to be the end of the world?
Either way, you don’t need to do it on every commit, just often enough that you catch these kinds of issues before they go to prod.
The speed and transparency of Cloudflare publishing this port mortem is excellent.
I also found the "remediation and follow up" section a bit lacking, not mentioning how, in general, regressions in query results caused by DB changes could be caught in future before they get widely rolled out.
Even if a staging env didn't have a production-like volume of data to trigger the same failure mode of a bot management system crash, there's also an opportunity to detect that something has gone awry if there were tests that the queries were returning functionally equivalent results after the proposed permission change. A dummy dataset containing a single http_requests_features column would suffice to trigger the dupe results behaviour.
In theory there's a few general ways this kind of issue could be detected, e.g. someone or something doing a before/after comparison to test that the DB permission change did not regress query results for common DB queries, for changes that are expected to not cause functional changes in behaviour.
Maybe it could have been detected with an automated test suite of the form "spin up a new DB, populate it with some curated toy dataset, then run a suite of important queries we must support and check the results are still equivalent (after normalising row order etc) to known good golden outputs". This style of regression testing is brittle, burdensome to maintain and error prone when you need to make functional changes and update what then "golden" outputs are - but it can give a pretty high probability of detecting that a DB change has caused unplanned functional regressions in query output, and you can find out about this in a dev environment or CI before a proposed DB change goes anywhere near production.
This wild `unwrap()` kinda took me aback as well. Someone really believed in themselves writing this. :)
They only recently rewrote their core in Rust (https://blog.cloudflare.com/20-percent-internet-upgrade/) -- given the newness of the system and things like "Over 100 engineers have worked on FL2, and we have over 130 modules" I won't be surprised for further similar incidents.
The irony of a rust rewrite taking down the internet is not lost on me.
I have to wonder if AI was involved with the change.
I don't think this is the case with CloudFlare, but for every recent GitHub outage or performance issue... oh boy, I blame the clankers!
hope no one was fired
So they made a newbie mistake in SQL that would not even pass an AI review. They did not verify the change in a test environment. And I guess the logs are so full of errors it is hard to pinpoint which matters. Yikes.
The internet hasn't been the internet in years. It was originally built to withstand wars. The whole idea of our IP based internet was to reroute packages should networks go down. Decentralisation was the mantra and how it differed from early centralised systems such as AOL et al.
This is all gone. The internet is a centralised system in the hand of just a few companies. If AWS goes down half the internet does. If Azure, Google Cloud, Oracle Cloud, Tencent Cloud or Alibaba Cloud goes down a large part of the internet does.
Yesterday with Cloudflare down half the sites I tried gave me nothing but errors.
The internet is dead.
It's not that deep, if AWS or Cloudflare suddenly disappeared sites would move to different hosts, it wouldn't mean the internet would die.
I mean you still depend on authoritative dns servers no?
Glad it's fixed, keep going!
this is where change management really shines because in a change management environment this would have been prevented by a backout procedure and it would never have been rolled out to production before going into QA, with peer review happening before that... I don't know if they lack change management but it's definitely something to think about
i think that is data rather than code which is where it falls short, in a way you need stringent code and more safeguarded code; it's like if everyone sends you 64k posts as that's all your proxy layer lets in, someone checked sending 128kb and it gave an error before reaching your app - and then someone sends 128kb and the proxy layer has changed - and your app crashes as it was more than 64kb and your app had an assert against that. to actually track issues with erraneous data that overflows well and stuff isn't so much code test but more like fuzz testing, brute force testing etc. which i think people should do; but that's more like we need strong test networks, and also those test networks may need to be more internet like to reflect real issues too, so the whole testing infrastructure in itself becomes difficult to get right - like they have their own tunneling system etc, they could segregate some of their servers and make a test system with better error diagnosis etc potentially. but to my mind, if they had better error propogation back that really identified what was happening and where then that would be a lot better in general. sure, start doing that on a test network. this is something i've beeen tihnking about in general - i made a simple rpc system for being able to send real time rust tracing logs (it allows to just use the normal tracing framework and use a thin rpc layer) back from multiple end servers but that's mostly for granular debugging. i've never quite understood why systems like systemd-journald aren't more network centric when they're going to be big and complex kitchensink approaches - apparently there's dbus support, but to my mind something inbetween debugging level of code and warning/info. like even if it's doing things like 1/20 of log info it's too much volume if things like large files getting close to limits is increasing etc and we can see this as things run, and can see if it's localised or common etc it'd help have more resilient systems. something may already exist in this line but i didn't come across anything in a reasonably passive way - i mean there's debugging tools like dtrace etc that have been around for ages.
It's unbelievable that the end of this postmortem is an advertisement for Cloudflare.
The last thing we need here is for more of the internet to sign up for Cloudflare.
It's in Cloudflare's interest and they wrote the blog post.
28M 500 errors/sec for several hours from a single provider. Must be a new record.
No other time in history has one single company been responsible for so much commerce and traffic. I wonder what some outage analogs to the pre-internet ages would be.
Something like a major telco going out, for example the AT&T 1990 outage of long distance calling:
> The standard procedures the managers tried first failed to bring the network back up to speed and for nine hours, while engineers raced to stabilize the network, almost 50% of the calls placed through AT&T failed to go through.
> Until 11:30pm, when network loads were low enough to allow the system to stabilize, AT&T alone lost more than $60 million in unconnected calls.
> Still unknown is the amount of business lost by airline reservations systems, hotels, rental car agencies and other businesses that relied on the telephone network.
https://users.csc.calpoly.edu/~jdalbey/SWE/Papers/att_collap...
Yes, all(most) eggs should not be in one basket. Perfect opportunity to setup a service that checks cloudflare then switches a site's DNS to akami as a backup.
> I wonder what some outage analogs to the pre-internet ages would be.
Lots of things have the sky in common. Maybe comet-induced ice ages...
Absolute volume maybe[1], as relative % of global digital communication traffic, the era of early telegraph probably has it beat.
In the pre digital era, East India Company dwarfs every other company in any metric like commerce controlled, global shipping, communication traffic, private army size, %GDP , % of workforce employed by considerable margins.
The default was large consolidated organization throughout history, like say Bell Labs, or Standard Oil before that and so on, only for a brief periods we have enjoyed benefits of true capitalism.
[1] Although I suspect either AWS or MS/Azure recent down-times in the last couple of years are likely higher
> No other time in history has one single company been responsible for so much commerce and traffic.
AWS very likely has Cloudflare beat in commerce responsibility. Amazon is equal to ~2.3% of US GDP by itself.
No publicity is bad publicity.
Best post mortem I've read in a while, this thing will be studied for years.
A bit ironic that their internal FL2 tool is supposed to make Cloudflare "faster and more secure" but brought a lot of things down. And yeah, as other have already pointed out, that's a very unsafe use of Rust, should've never made it to production.
Another long day...
But but but muh rust makes EVERYTHING safer!!!!
My dude, everything is a footgun if you hold it wrong enough
Great write up.
This is the first significant outage that has involved Rust code, and as you can see the .unwrap is known to carry the risk of a panic and should never be used on production code.
I give them a pass on lots of things, but this is inexcusable
I think you should give me a credit for all the income I lost due to this outage. Who authorized a change to the core infrastructure during the period of the year when your customers make the most income? Seriously, this is a management failure at the highest levels of decision-making. We don't make any changes to our server infrastructure/stack during the busiest time of the year, and neither should you. If there were an alternative to Cloudflare, I'd leave your service and move my systems elsewhere.
I think you should get exactly what the contract you signed said you'd get. Outages happen in all infrasturture. Planned and unplanned ones both. The SLA and SLO are literal acknowledgements of the fact that and part of the contract for that reason.
Its fair to be upset at their decision making - use that to renegotiate your contract.
This post was written by chatgpt????
https://blog.cloudflare.com/18-november-2025-outage/#:~:text...
Not all em dashes are ChatGPT. Good writers use it wherever required.
Did some $300k chief of IT blame it all on some overworked secretary clicking a link in an email they should have run through a filter? Because that’s the MO.
Hopefully that's the case.
Here is a bit more context in addition to the quote above. A ClickHouse permissions change made a metadata query start returning duplicate column metadata from an extra schema, which more than doubled the size and feature count of a Bot Management configuration file. When this oversized feature file was deployed to edge proxies, it exceeded a 200-feature limit in the bot module, causing that module to panic and the core proxy to return 5xx errors globally
So, to recap:
In terms of preventing this from a software engineering perspective, they made assumptions about how their database queries work (and didn't validate the results), and they ignored their own application limits and didn't program in either a test for whether an input would hit a limit, or some kind of alarm to notify the engineers of the source of the problem.From an operations perspective, it would appear they didn't test this on a non-production system mimicing production; they then didn't have a progressive deployment; and they didn't have a circuit breaker to stop the deployment or roll-back when a newly deployed app started crashing.
People jump to say things like "where's the rollback" and, like, probably yeah, but keep in mind that speculative rollback features (that is: rollbacks built before you've experienced the real error modes of the system) are themselves sources of sometimes-metastable distributed system failures. None of this is easy.
How about where's the most basic test to check if your config file will actually run at all in your application? It was a hard-coded memory limit; a git-hook test suite run a MacBook would have caught this. But nooo, let's not run the app for 0.01 seconds with this config before sending it out to determine the fate of the internet?
This is literally the CrowdStrike bug, in a CDN. This is the most basic, elementary, day 0 test you could possibly invent. Forget the other things they fucked up. Their app just crashes with a config file, and nobody evaluates it?! Not every bug is preventable, but an egregious lack of testing is preventable.
This is what a software building code (like the electrical code's UL listings that prevent your house from burning down from untested electrical components) is intended to prevent. No critical infrastructure should be legal without testing, period.
Looks like you have the perfect window to disrupt them with a superior product.
just before this outage i was exploring bunnycdn as the idea of cloudflare taking over dns still irks me slightly. there are competitors. but there's a certain amount of scale that cloudflare offers which i think can help performance in general. that said in the past i found cloudflare performance terrible when i was doings lots of testing. they are predominantly a pull based system not a push, so if content isn't current the cache miss performance can be kind of blah. i think their general backhaul paths have improved, but at least from new zealand they used to seem to do worse than hitting a los angeles proxy that then hits origin. (although google was in a similar position before, where both 8.8.8.8 and www.google.co.nz/.com were both faster via los angeles than via normal paths - i think google were doing asia parent, like if testing 8.8.8.8 misses it was super far away). i think now that we have http/3 etc though that performance is a bit simpler to achieve, and that ddos, bot protection is kind of the differentiator, and i think that cloudflare's bot protection may work reasonably well in general?