> The reason DBs like Mongo or Dynamo exist is because Postgres has a scaling problem.
I've used Postgres at a few places and the #1 problem was always high availability, not scaling. One Postgres cluster could easily handle 100000 transactions per minute, but when a primary node went down it was a page and manually failing over to the spare then manually replacing the spare. The manual tooling was very finicky but at least it worked, no automated solution came even close. Lack of a good HA story is why I avoid self-managed Postgres as much as possible.
I've extensively used Dynamo (internally at Amazon and externally) and even founded a DB startup with it at it's core. Boiling down scalability of Postgres vs Dynamo as it's written in blog is a bit terse. Dynamo scales writes horizontally with the keyspace, forever. Postgres simply can't, and no number of layers between the machines and the developer changes that. Sharding, pooling, Citus are all layered on top of an engine where a given row's writes still land on one
primary.
Curious how the DB startup with Dynamo at its core went. We use it heavily. The primary tricky thing for us at the moment is aligning pricing with workload value.
We obsessed over optimizations and pushing the apis to the limits of how we could pack it.
So much so, we re-wrote the DynamoSDK to squeeze out more optimizations so we could be the same cost even though we were a layer in front of dynamo. We used key encoding and other various technique as well as managed capacity (on demand vs reserved) to transparently optimize workloads for price. In our experience we saw dramatic gains vs just vanilla SDK usage.
If you're curious, here was the marketing website, but we're now part of Databricks: https://stately.cloud/
Interesting! We interact with the low-level APIs too vs the SDK, also: an IO scheduler for request batching and conn management, request hedging, full MVCC transactions, etc. We store raw bytes in DDB and manage schema/etc elsewhere. Curious if there is other low-hanging fruit, or not so low, you found that we haven't discovered yet.
Dynamo is a fundamentally different DB to Postgres. If your problem fits into the dynamo approach (I'd argue that more problems do), then you should be using it. No all problems fit, though.
Agreed, my critique was about how the article frames scalability. I've yet to see an OLTP problem that can't live in something like Dynamo. KV can model anything if you put in the work, the question is how much modeling discipline you trade for the scale, and in my experience the up front work is always worth it. Most of the time operational issues are swept under the rug and not consider tech debt.
Take for example AuroraDB: the sheer engineering it took to make SQL do scalable OLTP at all tells you how much that flexibility actually costs to keep.
Not by itself if it's naive, but if it's able to assess target health and avoid degraded instances then it becomes a component in HA, the other being integrating an orchestrator for gracious recovery.
> PgDog does not detect primary failure and will not call pg_promote(). It is expected that the databases are managed externally by another tool, like Patroni or AWS RDS, which handle replica promotion.
look, you guys can use a database tool from a guy who thinks he supports replication when the LLM he used to do everything correctly admitted that it doesn't support replication. i won't.
CNPG is quite nice and robust but I'd still be a bit reluctant to stack PG on k8s for really big clusters just because k8s ecosystem moves quite quickly and there's lots of patching/maintenance/churn which means more PG failovers so depends on how well your workload handles that (they're normally only a few seconds)
"Why Us" => "I ran Postgres at Instacart, where we scaled the company 5x in April of 2020. The biggest problem we had was making Postgres serve 100,000s of grocery delivery orders per minute"
That doesn't necessarily mean _new_ orders per minute. Their app or website could poll for updates every 15 seconds
Could just be looking at the "orders" endpoint in their app which might also include incremental updates as shoppers get items from the store. It's a fairly ambiguous statement
Instacard have released a public dataset[1] on their orders, so it should be even easier to verify this claim. From what I could find in some analysis[2] of this dataset around 100k orders per day and not per minute seems accurate.
I assume they are referring to how many database requests they have due to customers orders or a similar metric and just worded it poorly.
Is that still a lot? Feels like a single 64-core, 256GB RDS instance with some caching should handle that fine. RDS has instances up to 192-core and 768GB.
I am trying to gain a basic understanding of this:
Right now I have a 4TB DB on one large box.
Is the idea that using a proxy tool like PGDog I could spin up 8 smaller boxes handling ~500GB each and then one medium box for the proxy?
Right now I have a project that has very heavy write traffic from multiple services and a web app that reads from this.
We are starting to hit the point where no amount of indexing, query optimisation, caching or box upgrades is helping us.
We are looking at maybe moving the bulk of the static data to clickhouse to reduce the DB size but I would love to hear if PgDog or other kind of sharding could be useful for this use case.
> 8 smaller boxes handling ~500GB each and then one medium box for the proxy?
That's exactly right. Get in touch (lev@pgdog.dev), happy to help or at the very least tell you what current works (or doesn't) so you know what your options are.
I'm curious how this might help with our biggest downtime-causer with postgres, which is major version upgrades. Poolers do a great job for failover and load balancing, but we consistently need ~10-20 minutes of downtime once or twice a year to do upgrades. Logical replication between old->new versions could probably help, but it would still require flipping everything over to the new cluster without partial writes or anything silly. Anybody have experience with this?
Logical replication is how this is typically done. If you have some infra-as-code setup, you create a new cluster with identical settings except for the major version, import the schema, start copying data from a read-replica running the old version, stop accepting writes from the old version (downtime starts), sync the sequence numbers, and point your services to the new cluster (downtime ends).
If you use something like CloudNativePG they automate parts of the process with cli tools and declarative syntax. Otherwise you take the time to figure it out by hand. It might sound complicated, but just practice on your staging DB, and if all goes well you do the same procedure in prod.
Seconded. Coming from MySQL this is a huge regression that makes Postgres look like something from the 80s. I still wonder why this isn't seen as the absolutely highest priority.
I have not ran MySQL for some years but it at least used to have exactly the same issue. Upgrading a database with MySQL can take a long time if you have many tables. The main difference is only really that PostgreSQL does it with a separate tool, pg_upgrade, while MySQL does it as part of the main binary.
For both MySQL and PostgreSQL you will need to use some kind of logical upgrades if you want no downtime.
MySQL has advocated for decades spinning up a replica with the upgraded version, waiting for it to catch up to master before promoting it to the new master. You can do the same thing with Postgres.
Exactly, MySQL and PostgreSQL are the same here. Maybe one is a bit faster than the other at doing major version upgrades but the behaviours are quite similar.
It is also a bit tricky tradeoff. You do not want to be stuck with the same data format forever. So databases like MySQL and PostgreSQL need a downtime when doing a major version upgrade. They both try to keep it short, usually seconds, but minutes can happen in either database.
Logical replication needs a special 'upgrade' use case that will automate most of its pain points away. I understand why DDL does not replicate, and that you may want to replicate to a data warehouse that only needs some columns, etc, but there should be a case just for upgrading that handles all DDL, sequences all existing everything, and just works...
Do other RDBMSs have this? I genuinely have no clue. I've been fortunate enough to be able to get away with one primary and multiple secondaries at my largest usage of Postgres. Multi-master is the kind of thing I am fully out of my depth on, so I'm curious if there's a well defined path for implementation here or what.
Commercial RDBMS (oracle/mssql) have had it in some form for awhile, with pluses and minuses. Open source DBs have had bolt-ons, including BDR for pgsql.
Multi-master is hard. The main issue is what to do with commit/replication lag. It's far "easier" if support for eventual consistency is ok with your use case. In some cases it's not. Also, the problems related to read-only lag can happen on multi-master instances. If somebody does a giant long running query on one of the masters, the target instance needs to hold the data state for the query, even if the underlying DB is getting updates. It also needs to still keep up with other masters. This means the whole cluster can slow down if the multi-master replication is synchronous. Depending on a variety of factors, that can chew up disk space, memory, etc.
There are ways of dealing with these issues (and others), but it comes with tradeoffs with performance, etc.
It has been tried many times. Good luck to pgdog, but there’s a reason these projects don’t stick.
Multi master, from even a conceptual perspective, is incredibly complicated. Databases, transactions, consistency, parallelism are all very complicated.
It’s something that always seems promising at the start but as soon as maintenance and long term improvements enter the picture(ie integrating new Postgres versions), the complexity becomes too much.
If your working set is 20 TB, then it's pretty big. Each database has its own mix of hot/cold data, so it's impossible to compare without more information. A better measure might be IOPS. RDS has fairly low maximum IOPS unless you spend a lot more for provisioned IOPS or use Aurora.
You are correct. As a point of comparison: almost ten years ago at Segment we had a single Aurora PostgreSQL instance with ~50T of data, it was used to index potential identity data in a much larger corpus of files stored in S3.
I notice there is an Enterprise Edition, can you please specify which features are not open source? Do you predict new features you add will be ee licensed as a way to pay back your VC funders?
Is there an explainer for people who are broadly familiar with the DB space? It sounds like you're building an equivalent to Vitesse for Postgres, but it's not super clear from the article (which I know is not the point of this, but still :) ).
Edit: It also might be interesting to point out how your solution differs from what the folks at Planetscale are building https://planetscale.com/neki
I tried out PgDog a while ago, but couldn't find a good way of handling the config except for having this users / pgdog toml file, which makes it a bit awkward to handle in kubernetes where we often do multi-tenancy in postgres - or rather having many databases on the same instance(s), and have them come and go at will.
Also had an issue with it because it cached authentication requests when doing passthrough it seems, I'd changed the roles password, but it kept using the old one, which was no bueno ;).
PgDog seems to make more sense when you really care about a few databases that need massive scale, rather than a simple proxy in front of postgres. I'll keep following the development though, it is much needed in this space, postgres can use all the investment it can get to get it past the single machine scale that it excels at currently.
We successfully did this with pgdog at $JOB using our own "controller" -- the same service that handles deploying new instances of our application (instancing an argoCD Application that fires Crossplane DB creation, making new Deployments of bricks, etc) will also, at the end of that process, scan the cluster for Database CRDs, use those to generate a new pgdog.toml + users.toml, update the Secrets in the cluster, enable maintenance mode on all pgdog pods, do a live config reload on each of them, then disable maintenance mode (this is to make the change atomic between all the pgdog instances). Downtime there is about 2-3 seconds and all it does is make new SQL requests from existing clients wait, it doesn't break the connection or anything.
Not the place and not the time, but we are building an enterprise edition that "just works" out of the box. Not saying that the open source experience cannot be better - it always can and we'll keep improving. What you've experienced is definitely a known issue with our specific implementation of passthrough auth. Scram made things a bit harder, since we can't validate user's passwords at login time anymore (that's what makes scram secure fwiw).
Happy to chat about this, but we use the AWS secrets manager flowing into External Secrets Operator to generate a pgdog_users.toml. We then kick off a workflow to refresh things, but our rate of change here is much smaller than a super dynamic multi-tenant system.
You could also build a watcher side car that watches for changes of the pgdog_users.toml and have pgdog refresh itself then too with this combination. We thought about that but prefer to control the reloads for our needs.
> With $5.5M from Basis Set, YC, Pioneer Fund and other great investors, we have years of runway,
This is years of product development with a three person team. If Enterprise sales and support are a big part of your business plan it will suck up a lot more than that.
This is not an extension, it's a proxy! Very different. You can deploy it anywhere already without having to wait for upstreaming or your cloud provider adding support for it. It's one of the two reasons why we built it this way, the other being performance (it's much faster to do this in the proxy than inside Postgres).
> The reason DBs like Mongo or Dynamo exist is because
Not quite. The reason "DBs" like those exist is purely due to fashion. Lets not kid ourselves into thinking they do anything better, save the exception of making data hard to access, which might be a project goal in some cases.
>PgDog is a sharder, connection pooler and load balancer for PostgreSQL. Written in Rust, PgDog is fast, reliable and scales databases horizontally without requiring changes to application code.
Still trying to figure out how this works technically, is the performance gain really just re-write in rust?
Not quite. The performance gain is to bring those features to Postgres!
Edit:
Performance gains are from having the ability to load balance reads (horizontal scaling for read queries) and scale out writes (with sharding). Once instance bottleneck in Postgres has many faces:
1. Behind schedule vacuums because of too many dead tuples (too many writes)
2. The WALWriter is single-threaded and IO-bound - Postgres can only do about 200-300MB/sec in writes per instance (real prod numbers on EC2 with NVMes and ZFS, basically best case scenario).
3. Bulkheading: single primary is a single point of failure. With 12 primaries, if one fails, 91% of your customers don't notice.
The list goes on. Rust is just a side effect. We love it because it's fast and correct - the perfect match for a database product.
OLAP means different things to different people. For us, it's just making sure your admin dashboard keeps working basically:
SELECT tenant_id, COUNT(clicks)
FROM users
GROUP BY tenant_id
ORDER BY 2 DESC
LIMIT 25;
Performance is a side effect - definitely needed and we'll do everything we can, but we are not competing with ClickHouse or Snowflake - just trying to make sharded Postgres work with your app.
Re OLAP: It's probably ~good enough~ for a lean team that's trying to keep the tech stack standard and/or doesn't have a dedicated data person to take advantage of a columnar store.
I do tenant per PG schema, most are smallish some are bigger (not much, can do all in a single box) but moving forward eventually will need something like this. Also plan to provide "get your own VPS" for more enterprise customers.
I'm a big PGDog fan! It really helped us scale our connection proxy needs pretty substantially and it has great features like auto mode to support Aurora failovers neatly. It's infra that just works.
I've loved using pgdog for the last 6 months. It's been incredibly stable. It's nifty how they've solved the LISTEN/NOTIFY on a transaction pooler problem.
It’s surprising they don’t mention advantages over other sharding systems like Citus. Maybe it’s just the fact that it’s only a proxy and not core extensions? But that could limit capabilities.
The same old processes vs. threads debate, plus having the ability to scale the coordinator past a single machine. So, if you're OLTP, definitely consider PgDog. OLAP - Citus still wins because of its advanced query engine. We'll get there.
the reason mongo is a joy to use in scaled env is because no additional setup/software needed and all drivers natively support secondary/primary writes/reads and topological changes. so it's end to end, and adding is as a new proxy in frontend of postgres leads to all clients being incompatible or the code itself has no control anymore about when to use a secondary and what allowed stall is acceptable for a particular query. Any solutions to this by pgdog?
once mongo rewrote their engine - it's performant, scales & easy to run. seems a lot of devs got burnt by the early issues don't consider it all together.
its probably the easiest database to run at scale. run & forget. you just have to do a little more work on the data modeling part before you write your application i.e consider your query patterns.
> all drivers natively support secondary/primary writes/reads and topological changes.
Expanding on that a bit, mongo drivers even have a shared specification of the state machine for monitoring topology changes[1] and algorithm for selecting the server to send an operation to[2] (along with various declarative test cases that the drivers use to validate them alongside the specs in the repo). I think people sometimes underestimate how important the client-side work is to this sort of experience; for all of the faults mongo has had over the years, the amount of investment that they put into the client libraries is something I've never seen anywhere else (although having spent several years working on some of these libraries, my take is likely very biased).
Suggestion: have more than just helm and Docker in your quickstart documentation. I'd like to try this out just to see what it can do, but not quite enough to fire up one of those systems for it.
The docker compose example is just a demo. I don't know anyone who runs Postgres with docker compose / swarm in prod :) But yes, happy to add volumes so it seems more real.
We should add it to brew/apt/etc for sure. Also, we could add it to crates.io so you could do something like `cargo install pgdog`. Distribution, distribution, distribution.
Nit-Pick: It might be anti-marketing, still it would be helpful if the use cases can be articulated in a way where it would make sense to use this Vs any other type of database. Honesty goes a long way with the more technical folks for anything related to infrastructure.
Surfacing where and how PG is better than Dynamo or any other database is probably a good starting point instead of calling out PG a silver bullet for everything. At the end of the day its all a trade-off.
Depends. Only pooling, very little. Load balancing/sharding needs to parse queries, so a bit more. Could go up to a GB per pod, sometimes more if you have a lot of unique SQL queries (unique by text, not by parameters). We cache query ASTs to avoid parsing them on each request - that's the bulk of memory usage.
Semi related question - I have always wondered, how do you tackle OOM issues at the proxy layer, i.e. let's say a particular SQL query requires proxy to fan out the query to multiple shards, which return a pretty large dataset. I'm assuming you would need to load this dataset in the ram to perform certain operations. What happens if the resulting dataset causes the proxy pod to go OOM?
i am not using any tool like pgbouncer and have not run into any issues so far. Is it even required these days? Have you guys tested your setup without these connection poolers/multiplexers?
They are not just some random 3 have decades of real db experience behind them. They also just got funded which gives them the ability to expand and stay longer in the game.
Wrt. the pooler, how do you compare with pgbouncer?
I'm interested because I have a postgres instance, low-traffic but still like ... tens of r(eads)ps. I was not running anything close to the machine limits but still added pgbouncer to improve performance and didn't see a noticeable difference. I was stress-testing the machine obv., I'm not talking about the 10 rps, lol.
For context, my numbers were something like 10k rps +/- 1k vanilla postgres and like 9k rps +/- 1k with pgbouncer in front of it. So ... slightly slower but big error bars so I wouldn't say for sure. I ended up not using pgbouncer as the benefit was immaterial.
Also yeah, in case you want to check it out, it's the db that backs this project: https://httpstate.com.
> The reason DBs like Mongo or Dynamo exist is because Postgres has a scaling problem.
I've used Postgres at a few places and the #1 problem was always high availability, not scaling. One Postgres cluster could easily handle 100000 transactions per minute, but when a primary node went down it was a page and manually failing over to the spare then manually replacing the spare. The manual tooling was very finicky but at least it worked, no automated solution came even close. Lack of a good HA story is why I avoid self-managed Postgres as much as possible.
Good thing we support HA as well: https://docs.pgdog.dev/features/load-balancer/
Load balancer with health checks and failover, works out of the box. :) Battle-tested at this point too, so could be worth a look.
I've extensively used Dynamo (internally at Amazon and externally) and even founded a DB startup with it at it's core. Boiling down scalability of Postgres vs Dynamo as it's written in blog is a bit terse. Dynamo scales writes horizontally with the keyspace, forever. Postgres simply can't, and no number of layers between the machines and the developer changes that. Sharding, pooling, Citus are all layered on top of an engine where a given row's writes still land on one primary.
Curious how the DB startup with Dynamo at its core went. We use it heavily. The primary tricky thing for us at the moment is aligning pricing with workload value.
We obsessed over optimizations and pushing the apis to the limits of how we could pack it.
So much so, we re-wrote the DynamoSDK to squeeze out more optimizations so we could be the same cost even though we were a layer in front of dynamo. We used key encoding and other various technique as well as managed capacity (on demand vs reserved) to transparently optimize workloads for price. In our experience we saw dramatic gains vs just vanilla SDK usage.
If you're curious, here was the marketing website, but we're now part of Databricks: https://stately.cloud/
Interesting! We interact with the low-level APIs too vs the SDK, also: an IO scheduler for request batching and conn management, request hedging, full MVCC transactions, etc. We store raw bytes in DDB and manage schema/etc elsewhere. Curious if there is other low-hanging fruit, or not so low, you found that we haven't discovered yet.
Dynamo is a fundamentally different DB to Postgres. If your problem fits into the dynamo approach (I'd argue that more problems do), then you should be using it. No all problems fit, though.
Agreed, my critique was about how the article frames scalability. I've yet to see an OLTP problem that can't live in something like Dynamo. KV can model anything if you put in the work, the question is how much modeling discipline you trade for the scale, and in my experience the up front work is always worth it. Most of the time operational issues are swept under the rug and not consider tech debt.
Take for example AuroraDB: the sheer engineering it took to make SQL do scalable OLTP at all tells you how much that flexibility actually costs to keep.
That's great news! I'll bookmark this in case I'm forced to manage Postgres again.
What do you use instead?
Is a load balancer HA?
Not by itself if it's naive, but if it's able to assess target health and avoid degraded instances then it becomes a component in HA, the other being integrating an orchestrator for gracious recovery.
from their docs:
> PgDog does not detect primary failure and will not call pg_promote(). It is expected that the databases are managed externally by another tool, like Patroni or AWS RDS, which handle replica promotion.
look, you guys can use a database tool from a guy who thinks he supports replication when the LLM he used to do everything correctly admitted that it doesn't support replication. i won't.
Why the snark comment? The PgDog project has been around for a while, it's not vibe coded.
Combined with a replication strategy and automated health checks, a load balancer could direct traffic to a healthy instance automatically.
What happens when the load balancer fails?
Patroni 1.0 was released in 2016, i.e ~10 years ago.
https://github.com/patroni/patroni
Yup Patroni handles automatic failures and cluster management quite well
Patroni serves this niche pretty well at this point.
Have you looked into things like CloudnativePG? https://cloudnative-pg.io/
CNPG is quite nice and robust but I'd still be a bit reluctant to stack PG on k8s for really big clusters just because k8s ecosystem moves quite quickly and there's lots of patching/maintenance/churn which means more PG failovers so depends on how well your workload handles that (they're normally only a few seconds)
"Why Us" => "I ran Postgres at Instacart, where we scaled the company 5x in April of 2020. The biggest problem we had was making Postgres serve 100,000s of grocery delivery orders per minute"
Couldn't be a better why us :)
Is 100k order per minute a lot? Even a single Postgres instance should serve that fine?
100k(s) orders per minute is several orders of magnitude more than realistic. Amazon does 20k orders per minute.
Instacart doesn't need "100,000s of grocery delivery orders per minute".
There must be some 0s added for the sake of the story.
That doesn't necessarily mean _new_ orders per minute. Their app or website could poll for updates every 15 seconds
Could just be looking at the "orders" endpoint in their app which might also include incremental updates as shoppers get items from the store. It's a fairly ambiguous statement
According their 2026 Q1 filing they do about 90 million orders per quarter which is about 12 orders per second, 720 orders per minute.
It might make 100k row level changes per minute, but that’s a different metric.
https://www.sec.gov/Archives/edgar/data/1579091/000157909126...
Instacard have released a public dataset[1] on their orders, so it should be even easier to verify this claim. From what I could find in some analysis[2] of this dataset around 100k orders per day and not per minute seems accurate.
I assume they are referring to how many database requests they have due to customers orders or a similar metric and just worded it poorly.
[1] https://www.kaggle.com/datasets/psparks/instacart-market-bas... [2] https://rstudio-pubs-static.s3.amazonaws.com/284199_5c498037...
it could be peak orders per second
Amazon does 20k peak, or 20k average? Website visitor peaks could easily be two orders of magnitude higher traffic than average for a few minutes.
One assumes they mean 100,000s (plural) concurrent users actively building carts
Is that still a lot? Feels like a single 64-core, 256GB RDS instance with some caching should handle that fine. RDS has instances up to 192-core and 768GB.
Average throughput is one thing, tail latency, quite another.
I’ve always found Instacart to be extremely slow with giant latencies. Of course I don’t know if that’s due to Postgres or some other design flaw…
why did we switch to per minute? A modern quality enterprise SSD can do 35K +/- legit fsyncs per second.
Legends
I am trying to gain a basic understanding of this: Right now I have a 4TB DB on one large box. Is the idea that using a proxy tool like PGDog I could spin up 8 smaller boxes handling ~500GB each and then one medium box for the proxy?
Right now I have a project that has very heavy write traffic from multiple services and a web app that reads from this. We are starting to hit the point where no amount of indexing, query optimisation, caching or box upgrades is helping us. We are looking at maybe moving the bulk of the static data to clickhouse to reduce the DB size but I would love to hear if PgDog or other kind of sharding could be useful for this use case.
> 8 smaller boxes handling ~500GB each and then one medium box for the proxy?
That's exactly right. Get in touch (lev@pgdog.dev), happy to help or at the very least tell you what current works (or doesn't) so you know what your options are.
Scratching my head. Wondering why I would reach for this over just running a Yugabyte cluster.
I'm curious how this might help with our biggest downtime-causer with postgres, which is major version upgrades. Poolers do a great job for failover and load balancing, but we consistently need ~10-20 minutes of downtime once or twice a year to do upgrades. Logical replication between old->new versions could probably help, but it would still require flipping everything over to the new cluster without partial writes or anything silly. Anybody have experience with this?
We use logical replication and a pause / swap in pgbouncer for ~5s of paused (but not failed) writes.
This is for DBs that are ~1-1.5TB but doesnt have a huge amount of churn/qps
Effectively what is described here https://www.pgedge.com/blog/always-online-or-bust-zero-downt...
Logical replication is how this is typically done. If you have some infra-as-code setup, you create a new cluster with identical settings except for the major version, import the schema, start copying data from a read-replica running the old version, stop accepting writes from the old version (downtime starts), sync the sequence numbers, and point your services to the new cluster (downtime ends).
If you use something like CloudNativePG they automate parts of the process with cli tools and declarative syntax. Otherwise you take the time to figure it out by hand. It might sound complicated, but just practice on your staging DB, and if all goes well you do the same procedure in prod.
Edit: Apparently Postgres 19 has a patch for one-shot logical replication of sequences! https://www.depesz.com/2025/11/11/waiting-for-postgresql-19-...
RDS has blue green deployments that can help. It was rough at first, though seems they worked out the kinks.
Seconded. Coming from MySQL this is a huge regression that makes Postgres look like something from the 80s. I still wonder why this isn't seen as the absolutely highest priority.
I have not ran MySQL for some years but it at least used to have exactly the same issue. Upgrading a database with MySQL can take a long time if you have many tables. The main difference is only really that PostgreSQL does it with a separate tool, pg_upgrade, while MySQL does it as part of the main binary.
For both MySQL and PostgreSQL you will need to use some kind of logical upgrades if you want no downtime.
They don't change the on-disk structure all the time though...
Mostly because MySQL development is slower.
MySQL has advocated for decades spinning up a replica with the upgraded version, waiting for it to catch up to master before promoting it to the new master. You can do the same thing with Postgres.
Exactly, MySQL and PostgreSQL are the same here. Maybe one is a bit faster than the other at doing major version upgrades but the behaviours are quite similar.
Probably because it's an open source project and apparently none of its users cared about this feature enough to develop it or fund it.
It is also a bit tricky tradeoff. You do not want to be stuck with the same data format forever. So databases like MySQL and PostgreSQL need a downtime when doing a major version upgrade. They both try to keep it short, usually seconds, but minutes can happen in either database.
Logical replication solves this. You roll the cluster, downtime is minimal. like 60s maybe.
Logical replication needs a special 'upgrade' use case that will automate most of its pain points away. I understand why DDL does not replicate, and that you may want to replicate to a data warehouse that only needs some columns, etc, but there should be a case just for upgrading that handles all DDL, sequences all existing everything, and just works...
It's weird that PostgreSQL still doesn't have a proper, open source, general multi-master implementation.
At this point i wonder if i'll ever see that.
Do other RDBMSs have this? I genuinely have no clue. I've been fortunate enough to be able to get away with one primary and multiple secondaries at my largest usage of Postgres. Multi-master is the kind of thing I am fully out of my depth on, so I'm curious if there's a well defined path for implementation here or what.
Commercial RDBMS (oracle/mssql) have had it in some form for awhile, with pluses and minuses. Open source DBs have had bolt-ons, including BDR for pgsql.
Multi-master is hard. The main issue is what to do with commit/replication lag. It's far "easier" if support for eventual consistency is ok with your use case. In some cases it's not. Also, the problems related to read-only lag can happen on multi-master instances. If somebody does a giant long running query on one of the masters, the target instance needs to hold the data state for the query, even if the underlying DB is getting updates. It also needs to still keep up with other masters. This means the whole cluster can slow down if the multi-master replication is synchronous. Depending on a variety of factors, that can chew up disk space, memory, etc.
There are ways of dealing with these issues (and others), but it comes with tradeoffs with performance, etc.
MySQL has Galera cluster for that.
And Group Replication
It has been tried many times. Good luck to pgdog, but there’s a reason these projects don’t stick.
Multi master, from even a conceptual perspective, is incredibly complicated. Databases, transactions, consistency, parallelism are all very complicated.
It’s something that always seems promising at the start but as soon as maintenance and long term improvements enter the picture(ie integrating new Postgres versions), the complexity becomes too much.
PgDog, Neki, multigres, awesome to see. And yes this is the main issue with postgres. Well this and not having index hints, looking forward to 19
The pg_hint_plan extension isn't in core, yet is pretty competent when you need to override planner.
Don't forget the original PgBouncer. Hard to setup, but with the help of AI these days it's easier to configure.
If your working set is 20 TB, then it's pretty big. Each database has its own mix of hot/cold data, so it's impossible to compare without more information. A better measure might be IOPS. RDS has fairly low maximum IOPS unless you spend a lot more for provisioned IOPS or use Aurora.
You are correct. As a point of comparison: almost ten years ago at Segment we had a single Aurora PostgreSQL instance with ~50T of data, it was used to index potential identity data in a much larger corpus of files stored in S3.
For a vast majority of use cases 20TB is positively enormous.
RDS caps out at 64 TB unless you use Aurora, so 20 TB is totally manageable without sharding.
This product is for Postgres deployments that are so large they need to be sharded. For these use cases, I think 20TB is about normal.
Yes. But for most workloads it is not much for PostgreSQL. You often will not have to shard at all.
Sure, but 20TB in “the only database you need” is mere hours or minutes worth of data for many workflows.
that article seems to suggest 20TB total over the dozen deployments in prod.
I notice there is an Enterprise Edition, can you please specify which features are not open source? Do you predict new features you add will be ee licensed as a way to pay back your VC funders?
Two big ones:
1. Control plane to manage multi-node deployments; "works out of the box" experience to make PgDog easy to deploy and use
2. QoS (quality of service): automatically block bad queries from taking down the database
Last but not least, you get SLA-backed support from us (up to P0).
New features are broken down into two categories:
1. Sharding / running Postgres at scale: always open source.
2. Infra management / making it easy to run PgDog at scale: enterprise.
Thank you for the clear response!
I us pg. not that I know much about database internals, besides the 'b-tree' stuff we learned in college.
I don't know how the pg scaling story gets fixed unless certain things are rewritten. that's my fear of going all in pg.
mysql has vitess etc & even upgrades are easier. though pg is more extensible.
Is there an explainer for people who are broadly familiar with the DB space? It sounds like you're building an equivalent to Vitesse for Postgres, but it's not super clear from the article (which I know is not the point of this, but still :) ).
Edit: It also might be interesting to point out how your solution differs from what the folks at Planetscale are building https://planetscale.com/neki
There's multiple solutions coming up in this space:
1. Neki as you mentioned 2. PgDog 3. Multigres, headed by original creator of Vitesse
I tried out PgDog a while ago, but couldn't find a good way of handling the config except for having this users / pgdog toml file, which makes it a bit awkward to handle in kubernetes where we often do multi-tenancy in postgres - or rather having many databases on the same instance(s), and have them come and go at will.
Also had an issue with it because it cached authentication requests when doing passthrough it seems, I'd changed the roles password, but it kept using the old one, which was no bueno ;).
PgDog seems to make more sense when you really care about a few databases that need massive scale, rather than a simple proxy in front of postgres. I'll keep following the development though, it is much needed in this space, postgres can use all the investment it can get to get it past the single machine scale that it excels at currently.
We successfully did this with pgdog at $JOB using our own "controller" -- the same service that handles deploying new instances of our application (instancing an argoCD Application that fires Crossplane DB creation, making new Deployments of bricks, etc) will also, at the end of that process, scan the cluster for Database CRDs, use those to generate a new pgdog.toml + users.toml, update the Secrets in the cluster, enable maintenance mode on all pgdog pods, do a live config reload on each of them, then disable maintenance mode (this is to make the change atomic between all the pgdog instances). Downtime there is about 2-3 seconds and all it does is make new SQL requests from existing clients wait, it doesn't break the connection or anything.
Not the place and not the time, but we are building an enterprise edition that "just works" out of the box. Not saying that the open source experience cannot be better - it always can and we'll keep improving. What you've experienced is definitely a known issue with our specific implementation of passthrough auth. Scram made things a bit harder, since we can't validate user's passwords at login time anymore (that's what makes scram secure fwiw).
We'll get there.
Happy to chat about this, but we use the AWS secrets manager flowing into External Secrets Operator to generate a pgdog_users.toml. We then kick off a workflow to refresh things, but our rate of change here is much smaller than a super dynamic multi-tenant system.
You could also build a watcher side car that watches for changes of the pgdog_users.toml and have pgdog refresh itself then too with this combination. We thought about that but prefer to control the reloads for our needs.
Three real-world issues I've run into recently with PgBouncer + Postgres are:
1. pool exhaustion from idle connections inside open long-running transactions
2. SQLAlchemy's client-side pool using dead connections that PgBouncer had already killed, causing periodic request errors
3. Some tasks have to bypass PgBouncer when they use SET or prepared statements
I've already sharded large datasets at the application layer, but looks like PgDog solves the above problems for any future work?
SQLA async is a bit of a struggle with pgbouncer.
I had to disable application pooling as it was causing read only transactions I could couldnt pin down the cause.
> With $5.5M from Basis Set, YC, Pioneer Fund and other great investors, we have years of runway,
This is years of product development with a three person team. If Enterprise sales and support are a big part of your business plan it will suck up a lot more than that.
Presumably enterprise sales will bring in revenue on its own
I've seen a couple of these "distributed" postgres extensions.
My question is, has any of them been talked about being upstreamed to postgres itself? Or, adding a custom built in feature to postgres itself?
This is not an extension, it's a proxy! Very different. You can deploy it anywhere already without having to wait for upstreaming or your cloud provider adding support for it. It's one of the two reasons why we built it this way, the other being performance (it's much faster to do this in the proxy than inside Postgres).
> The reason DBs like Mongo or Dynamo exist is because
Not quite. The reason "DBs" like those exist is purely due to fashion. Lets not kid ourselves into thinking they do anything better, save the exception of making data hard to access, which might be a project goal in some cases.
>PgDog is a sharder, connection pooler and load balancer for PostgreSQL. Written in Rust, PgDog is fast, reliable and scales databases horizontally without requiring changes to application code.
Still trying to figure out how this works technically, is the performance gain really just re-write in rust?
Not quite. The performance gain is to bring those features to Postgres!
Edit:
Performance gains are from having the ability to load balance reads (horizontal scaling for read queries) and scale out writes (with sharding). Once instance bottleneck in Postgres has many faces:
1. Behind schedule vacuums because of too many dead tuples (too many writes)
2. The WALWriter is single-threaded and IO-bound - Postgres can only do about 200-300MB/sec in writes per instance (real prod numbers on EC2 with NVMes and ZFS, basically best case scenario).
3. Bulkheading: single primary is a single point of failure. With 12 primaries, if one fails, 91% of your customers don't notice.
The list goes on. Rust is just a side effect. We love it because it's fast and correct - the perfect match for a database product.
So to oversimplify, is the idea to bring an AWS Aurora-style storage mechanism natively to Postgres?
Yes, except it doesn't have any cross-dependencies on the same volume, so the uptime here should be higher.
Aurora has a completely different storage backend. PgDog is a front end proxy - each server in the cluster is still using standard Postgres right?
Yup!
Oh thanks for clearing that up.
Sorry, out walking the dog (not a pun). I'll post more details in a few.
Good stuff, although I’m not quite sure about the fast OLAP use case.
If you’re already sharding by tenant for other reasons, OK… But I see CDC to a true OLAP system as more scalable.
PostgreSQL still needs real columnar tables in the core, hopefully one day
OLAP means different things to different people. For us, it's just making sure your admin dashboard keeps working basically:
Performance is a side effect - definitely needed and we'll do everything we can, but we are not competing with ClickHouse or Snowflake - just trying to make sharded Postgres work with your app.Re OLAP: It's probably ~good enough~ for a lean team that's trying to keep the tech stack standard and/or doesn't have a dedicated data person to take advantage of a columnar store.
I do tenant per PG schema, most are smallish some are bigger (not much, can do all in a single box) but moving forward eventually will need something like this. Also plan to provide "get your own VPS" for more enterprise customers.
This kind of tool will help in this case?
Yup. We support schema-based sharding: https://docs.pgdog.dev/configuration/pgdog.toml/sharded_sche...
I'm a big PGDog fan! It really helped us scale our connection proxy needs pretty substantially and it has great features like auto mode to support Aurora failovers neatly. It's infra that just works.
I've moved from pgbouncer to pgdog a few months ago without issue. Huge fan.
I've loved using pgdog for the last 6 months. It's been incredibly stable. It's nifty how they've solved the LISTEN/NOTIFY on a transaction pooler problem.
It’s surprising they don’t mention advantages over other sharding systems like Citus. Maybe it’s just the fact that it’s only a proxy and not core extensions? But that could limit capabilities.
We do, just buried deep in our blog: https://pgdog.dev/blog/pgdog-vs-citus
The same old processes vs. threads debate, plus having the ability to scale the coordinator past a single machine. So, if you're OLTP, definitely consider PgDog. OLAP - Citus still wins because of its advanced query engine. We'll get there.
the reason mongo is a joy to use in scaled env is because no additional setup/software needed and all drivers natively support secondary/primary writes/reads and topological changes. so it's end to end, and adding is as a new proxy in frontend of postgres leads to all clients being incompatible or the code itself has no control anymore about when to use a secondary and what allowed stall is acceptable for a particular query. Any solutions to this by pgdog?
once mongo rewrote their engine - it's performant, scales & easy to run. seems a lot of devs got burnt by the early issues don't consider it all together.
its probably the easiest database to run at scale. run & forget. you just have to do a little more work on the data modeling part before you write your application i.e consider your query patterns.
> all drivers natively support secondary/primary writes/reads and topological changes.
Expanding on that a bit, mongo drivers even have a shared specification of the state machine for monitoring topology changes[1] and algorithm for selecting the server to send an operation to[2] (along with various declarative test cases that the drivers use to validate them alongside the specs in the repo). I think people sometimes underestimate how important the client-side work is to this sort of experience; for all of the faults mongo has had over the years, the amount of investment that they put into the client libraries is something I've never seen anywhere else (although having spent several years working on some of these libraries, my take is likely very biased).
[1]: https://github.com/mongodb/specifications/blob/master/source... [2]: https://github.com/mongodb/specifications/blob/master/source...
How is this different from Citus?
I wish them all the best. Supabase, Timescale, etc etc. there's a whole cottage industry of extending postgres to whatever you need.
Suggestion: have more than just helm and Docker in your quickstart documentation. I'd like to try this out just to see what it can do, but not quite enough to fire up one of those systems for it.
Is there a binary I can run directly?
In addition - the docker compose example doesn't set up any data volumes for the postgres instances - that might be considered a bug?
Then again, sharding on a single host probably isn't very useful anyway - but it might work with docker in swarm mode?
The docker compose example is just a demo. I don't know anyone who runs Postgres with docker compose / swarm in prod :) But yes, happy to add volumes so it seems more real.
We should add it to brew/apt/etc for sure. Also, we could add it to crates.io so you could do something like `cargo install pgdog`. Distribution, distribution, distribution.
I also appreciate GitHub releases with pre-compiled binaries for different platforms. The more options the better!
Does making it "just work" here come with any caveats vs standard PG?
Getting there! Cross-shard writes do because of 2pc. Reads are eventually consistent.
Given that they implement connection pooling and sharding, I'm going to say "not at all".
You _could_ make that ACID, but it's not going to be faster than a single machine.
Nit-Pick: It might be anti-marketing, still it would be helpful if the use cases can be articulated in a way where it would make sense to use this Vs any other type of database. Honesty goes a long way with the more technical folks for anything related to infrastructure.
Surfacing where and how PG is better than Dynamo or any other database is probably a good starting point instead of calling out PG a silver bullet for everything. At the end of the day its all a trade-off.
Always is. Marketing is not our strong suit (only engineers here). We'll get better at it.
2M qps in production is legit. Curious how much RAM and CPU that takes on average per deployment though
Depends. Only pooling, very little. Load balancing/sharding needs to parse queries, so a bit more. Could go up to a GB per pod, sometimes more if you have a lot of unique SQL queries (unique by text, not by parameters). We cache query ASTs to avoid parsing them on each request - that's the bulk of memory usage.
Semi related question - I have always wondered, how do you tackle OOM issues at the proxy layer, i.e. let's say a particular SQL query requires proxy to fan out the query to multiple shards, which return a pretty large dataset. I'm assuming you would need to load this dataset in the ram to perform certain operations. What happens if the resulting dataset causes the proxy pod to go OOM?
Two schools of thought:
1. Let it crash. Increase the RAM, try again.
2. Page to disk (swap), make it slow but ultimately work.
Both have their trade-offs. There is no free lunch here.
I hope people pronounce this as „pig-dog” and has a mascot that looks like „man-bear-pig”
Crap! Missed opportunity.
i am not using any tool like pgbouncer and have not run into any issues so far. Is it even required these days? Have you guys tested your setup without these connection poolers/multiplexers?
Each connection is a process on the server, that takes up both CPU and RAM, it will run out.
This solves the thousands of clients case for read in a way that is transparent to the clients.
Yes it's required at large scale, especially if you want to distribute reads or shard to a particular geographical area.
This commit looks... odd.
https://github.com/pgdogdev/pgdog/commit/36434f93f03dec1d7d4...
I want to have as much fun as the next developer, but that makes me worry, what with supply chain attacks in the news and all.
I see you met Sage, our newest founding engineer :) If you're not having fun at your job...
In all seriousness, we review every single line of code that goes in and only people who work for PgDog Inc are allowed to merge.
How are 3 developers going to QA this properly ?
They are not just some random 3 have decades of real db experience behind them. They also just got funded which gives them the ability to expand and stay longer in the game.
How are 3 developers going to sell that to any company? Procurement will have a field day.
They have funding. That's what it will be for. I wish them well and appreciate that people are still doing FOSS.
As long as they don't get undercut by the equivalent of AWS https://aws.amazon.com/rds/proxy/ which is a managed pgbouncer.
The issue is if the DB layer fails your product is going to completely stop working.
You’d need a ton of faith in these 3 people.
Feels more like it would work better inside of a bigger organization.
The QA tester in me is kinda risk adverse.
we are using PG bouncer in production. Interesting, I will follow the evolution of this project
how does it compare to PlanetScale ?
PgDog is GalaxyScale </joke>
Cool work, thanks.
Wrt. the pooler, how do you compare with pgbouncer?
I'm interested because I have a postgres instance, low-traffic but still like ... tens of r(eads)ps. I was not running anything close to the machine limits but still added pgbouncer to improve performance and didn't see a noticeable difference. I was stress-testing the machine obv., I'm not talking about the 10 rps, lol.
For context, my numbers were something like 10k rps +/- 1k vanilla postgres and like 9k rps +/- 1k with pgbouncer in front of it. So ... slightly slower but big error bars so I wouldn't say for sure. I ended up not using pgbouncer as the benefit was immaterial.
Also yeah, in case you want to check it out, it's the db that backs this project: https://httpstate.com.
Old benchmark, but still good: https://pgdog.dev/blog/pgbouncer-vs-pgdog