One thing that bothered me enough to comment on: “transcoding” doesn’t seem like the right term in this context. “Encode/decode” is technically correct, but “compress/decompress” would have made the intent much clearer.
Tangentially related, I applied a similar approach to compress the npm registry by over 90% on disk a few years back. Since most versions of a package are similar, you can delta encode them first and then compress them. The deltas are small and compress well as a collection with the original source files.
For another use case, prior to compressing, I’ve applied a rolling hash to deterministically split the file. Then compressed the chunks and stored them in a CID filesystem. The result is that files that are largely similar share compressed chunks.
There are a lot of things we can do to be substantially more efficient with the computers we have, but engineers often cost more than hardware. With recent supply chain constraints that calculus is changing!
NPM packages doesn’t take up that much disk space for me with standard pnpm deduplication. The much bigger offender for me is Rust target dir: when people talk about vibe coding in Rust for performance, what they don’t mention (at least I’ve hardly ever seen it mentioned) is every trivial little tool pushes 1GB on disk and anything slightly nontrivial easily racks up multi-GB. Which hurts when you have lots of vibed small tools. I wonder if anyone bothered to look into this problem.
I think the parent was referring to hosting the entire NPM registry, not having a project that uses NPM. In that case there's substantial duplication from version to version.
I've done similar things for large container images. My format allows for using FastCDC to chunk files, but there's a tradeoff between number of shared chunks and between number of HTTP requests. I keep it turned off by default.
On the side I'm working on an extension to git-lfs to use fastcdc for both storage and transmission of artifacts to drastically reduce size and make git-lfs more practical for more things.
I'm confused by how this affects range requests. Without compression, those can be easily satisfied by reading the relevant part of the cached complete file. But how are they handled now? The article claims "range requests remain unchanged", but I don't see how that's possible if the cache no longer stores the uncompressed data.
I assume the entire resource needs to be decompressed first, then indexed into, served, and discarded. Well, actually, you could just decompress up to the end of the range.
Actually zstd internally splits data into frames, and frames can indicate the decompressed data size. So if we control the compressor we can make it so that all frames have the size information; it isn’t exactly seekable but at least it will not need to decompress the resource. https://python-zstandard.readthedocs.io/en/latest/concepts.h...
Given how fast zstd can decompress, this may or may not actually be a win: the time spent waiting for I/O might be so large that the decompression can fit within the wait time.
Which would have terrible performance for range requests starting late in a large file. For files that are frequently accessed that way, this could be prohibitive.
You could split the file into independently compressed blocks as well. But that'd reduce compression rate and require adding some kind of index for seeking.
Or they have an upper size limit for the file size they compress, since large files are rarely compressible text.
In any case it is something that needs the be handled before going live with a compressed cache. But the article sounds like they simply didn't implement compressed caching for those cases, which makes no sense.
For the cost of a small amount of metadata the offsets of every MiB or so could be stored. I did something like this with pigz as I was implementing multithreaded compressed and encrypted kernel zone suspend and resume for Solaris.
Not with zstd, you could still support range requests. https://en.wikipedia.org/wiki/Zstd this whole subthread should take 10 minutes and glance over the spec and the capabilities. It would end a lot of wasted premature pontificating.
Seekable OCI (SOCI) uses an index so I imagine that's an option (real byte range a-b maps to compressed range x-y). Presumably you'd still need to read the header and some additional pieces
ZFS compresses recordsize or volblocksize chunks down to some whole number of disk blocks, as determined by 1 >> ashift. In practice, this typically means that each 128k chunk gets compressed to some number of sequential 512 or 4096 byte blocks. These compressed blocks are referenced by block pointers that contain flags indicating compression and what type.
When I worked on a large CDN the content popularity distribution was heavily skewed. Think 20-40% of throughput from top 1% if content, and 80-90% from the top 10%. Anything outside of that had a very low probability if ever being read again in the effective cache lifetime.
Then the effective cost of scaling CPU > RAM > storage > network due to power & space limitations. Spending extra processing time on 50% of your content would be wasted effort as its never read again.
And yes, increasing effective storage might increase cache width/lifetime, but its not by that much to dramatically inprove access rates. Especially when most content by unique bytes is compressed audio/video in the first place.
I would compress it all, and then selectively recompress at higher compression levels depending on the link, read frequency, diversity and capabilities of the clients.
Zstd 3 to 5 is nearly free in terms of not bottlenecking disk or network. Zstd 12 to 19 gives amazing compression results and still result in speedups when reading from disk. It really is a wonderful all purpose compressor.
One of the nice things about Zstd is if you try to compress an already compressed stream, it short circuits. So even if you are given say HVEC MP4 and run zstd -19 on it, it will "compress" immediately and not DOS your pipeline.
I think they probably don't care about storage on the devices that do the compressing and are optimizing for quickly pushing hot content to edge locations. So the compression at the source saves bandwidth during the pushing to edge phase and allows the edges to hold more (reducing churn, further saving bandwidth back to the source).
Put a different way, they're trying to make cache evictions cheaper (less bandwidth to refill) and less likely (bigger cache on same disk size)
"Ignore cold assets" makes more sense with that framing
Although if that's the case, the CPU statement still is a bit confusing.
I agree! I came to the comment section to say exactly this. In any cache hierarchy you want to put colder content in cheaper but slower storage. Here, compression is the cheaper but slower form of storage.
It could save PyPI petabytes per month of bandwidth, too. (But it seems like this is also caused by broken CI systems failing to cache things locally.)
Yeah. It's worth looking at your own code and infrastructure as well. In my previous job there were dozens of opportunities to realise big savings by swapping out gzip for other compression schemes -- usually zstd, occasionally lz4 or bz2 or brotli. Huge assets took less hard drive space, took less time to download, and took less time to decompress. The differences were not small, and resulted in appreciable improvements both in infra cost and dev productivity.
Curious how this application scheme compares to filesystem + transport compression. You'd end up potentially compressing and decompressing more often but the higher software doesn't need to know what's happening and the compression happens in kernel space.
ie btrfs
You could also layer on out of band dedupe and probably push out cache updates with btrfs snapshots although maybe that ends too convoluted
One thing that bothered me enough to comment on: “transcoding” doesn’t seem like the right term in this context. “Encode/decode” is technically correct, but “compress/decompress” would have made the intent much clearer.
Tangentially related, I applied a similar approach to compress the npm registry by over 90% on disk a few years back. Since most versions of a package are similar, you can delta encode them first and then compress them. The deltas are small and compress well as a collection with the original source files.
For another use case, prior to compressing, I’ve applied a rolling hash to deterministically split the file. Then compressed the chunks and stored them in a CID filesystem. The result is that files that are largely similar share compressed chunks.
There are a lot of things we can do to be substantially more efficient with the computers we have, but engineers often cost more than hardware. With recent supply chain constraints that calculus is changing!
NPM packages doesn’t take up that much disk space for me with standard pnpm deduplication. The much bigger offender for me is Rust target dir: when people talk about vibe coding in Rust for performance, what they don’t mention (at least I’ve hardly ever seen it mentioned) is every trivial little tool pushes 1GB on disk and anything slightly nontrivial easily racks up multi-GB. Which hurts when you have lots of vibed small tools. I wonder if anyone bothered to look into this problem.
I think the parent was referring to hosting the entire NPM registry, not having a project that uses NPM. In that case there's substantial duplication from version to version.
I've done similar things for large container images. My format allows for using FastCDC to chunk files, but there's a tradeoff between number of shared chunks and between number of HTTP requests. I keep it turned off by default.
I'd be curious whether block level de-duping would add value too in their case. You effectively achieved that to some degree with storing the deltas.
You may or may not be familiar with the content defined chunking family of algorithms like FastCDC.
https://joshleeb.com/posts/chunking.html
https://www.usenix.org/conference/atc16/technical-sessions/p...
On the side I'm working on an extension to git-lfs to use fastcdc for both storage and transmission of artifacts to drastically reduce size and make git-lfs more practical for more things.
Why not serving files compressed if the client supports it even though the origin served an uncompressed file?
I was thinking this. Zstd is widely supported in browsers, over 80% right now and will increase over time:
https://caniuse.com/?search=zstd
They already do this, at least for paid accounts. You can even decide what compression model you want them to serve on your behalf.
I'm confused by how this affects range requests. Without compression, those can be easily satisfied by reading the relevant part of the cached complete file. But how are they handled now? The article claims "range requests remain unchanged", but I don't see how that's possible if the cache no longer stores the uncompressed data.
> I don't see how that's possible if the cache no longer stores the uncompressed data.
Zstd has a seekable format for frames, similar to pigz --independent works.
[1] - https://github.com/facebook/zstd/blob/dev/contrib/seekable_f...
Yes. This.
I assume the entire resource needs to be decompressed first, then indexed into, served, and discarded. Well, actually, you could just decompress up to the end of the range.
Actually zstd internally splits data into frames, and frames can indicate the decompressed data size. So if we control the compressor we can make it so that all frames have the size information; it isn’t exactly seekable but at least it will not need to decompress the resource. https://python-zstandard.readthedocs.io/en/latest/concepts.h...
Given how fast zstd can decompress, this may or may not actually be a win: the time spent waiting for I/O might be so large that the decompression can fit within the wait time.
Which would have terrible performance for range requests starting late in a large file. For files that are frequently accessed that way, this could be prohibitive.
You could split the file into independently compressed blocks as well. But that'd reduce compression rate and require adding some kind of index for seeking.
Or they have an upper size limit for the file size they compress, since large files are rarely compressible text.
In any case it is something that needs the be handled before going live with a compressed cache. But the article sounds like they simply didn't implement compressed caching for those cases, which makes no sense.
For the cost of a small amount of metadata the offsets of every MiB or so could be stored. I did something like this with pigz as I was implementing multithreaded compressed and encrypted kernel zone suspend and resume for Solaris.
Not with zstd, you could still support range requests. https://en.wikipedia.org/wiki/Zstd this whole subthread should take 10 minutes and glance over the spec and the capabilities. It would end a lot of wasted premature pontificating.
There's nothing about random access at that link.
There's this, but it doesn't seem to be getting much traction: https://github.com/facebook/zstd/tree/dev/contrib/seekable_f...
Also external dictionaries https://nigeltao.github.io/blog/2022/zstandard-part-7-dictio...
Idk but btrfs and zfs manage to pull it off
Seekable OCI (SOCI) uses an index so I imagine that's an option (real byte range a-b maps to compressed range x-y). Presumably you'd still need to read the header and some additional pieces
ZFS compresses recordsize or volblocksize chunks down to some whole number of disk blocks, as determined by 1 >> ashift. In practice, this typically means that each 128k chunk gets compressed to some number of sequential 512 or 4096 byte blocks. These compressed blocks are referenced by block pointers that contain flags indicating compression and what type.
afaik, ZFS will read an entire Record at a time- and that's the same granularity as its compression.
It uses some form of keyframing. Not entirely free.
they said they didn't change the behavior for range requests. So, it'll still be the basic no compression. (eg, server side decompression)
> We initially considered limiting transcoding to popular content
Weird, I would have compressed cold content instead, if the goal was to save on CPU time during decode.
When I worked on a large CDN the content popularity distribution was heavily skewed. Think 20-40% of throughput from top 1% if content, and 80-90% from the top 10%. Anything outside of that had a very low probability if ever being read again in the effective cache lifetime.
Then the effective cost of scaling CPU > RAM > storage > network due to power & space limitations. Spending extra processing time on 50% of your content would be wasted effort as its never read again.
And yes, increasing effective storage might increase cache width/lifetime, but its not by that much to dramatically inprove access rates. Especially when most content by unique bytes is compressed audio/video in the first place.
I would compress it all, and then selectively recompress at higher compression levels depending on the link, read frequency, diversity and capabilities of the clients.
Zstd 3 to 5 is nearly free in terms of not bottlenecking disk or network. Zstd 12 to 19 gives amazing compression results and still result in speedups when reading from disk. It really is a wonderful all purpose compressor.
One of the nice things about Zstd is if you try to compress an already compressed stream, it short circuits. So even if you are given say HVEC MP4 and run zstd -19 on it, it will "compress" immediately and not DOS your pipeline.
It has mostly become my go-to as well. Just wish it wasn't a Facebook product.
That part is a little bit confusing.
I think they probably don't care about storage on the devices that do the compressing and are optimizing for quickly pushing hot content to edge locations. So the compression at the source saves bandwidth during the pushing to edge phase and allows the edges to hold more (reducing churn, further saving bandwidth back to the source).
Put a different way, they're trying to make cache evictions cheaper (less bandwidth to refill) and less likely (bigger cache on same disk size)
"Ignore cold assets" makes more sense with that framing
Although if that's the case, the CPU statement still is a bit confusing.
I agree! I came to the comment section to say exactly this. In any cache hierarchy you want to put colder content in cheaper but slower storage. Here, compression is the cheaper but slower form of storage.
It could save PyPI petabytes per month of bandwidth, too. (But it seems like this is also caused by broken CI systems failing to cache things locally.)
Yeah. It's worth looking at your own code and infrastructure as well. In my previous job there were dozens of opportunities to realise big savings by swapping out gzip for other compression schemes -- usually zstd, occasionally lz4 or bz2 or brotli. Huge assets took less hard drive space, took less time to download, and took less time to decompress. The differences were not small, and resulted in appreciable improvements both in infra cost and dev productivity.
Curious how this application scheme compares to filesystem + transport compression. You'd end up potentially compressing and decompressing more often but the higher software doesn't need to know what's happening and the compression happens in kernel space.
ie btrfs
You could also layer on out of band dedupe and probably push out cache updates with btrfs snapshots although maybe that ends too convoluted