I love seeing stuff like this that would probably not exist if not for LLMs making exploring these kinds of ideas relatively cheap and quick to do.
My takeaway from this though is that nginx is pretty impressive on its own. Also this stuck out:
It's meant to be an alternative to nginx and Caddy, and the design bet is about configuration. Those servers give you a declarative config language - location blocks, rewrite rules, map directives, try_files - and then, once the declarative language hits its limits, an optional scripting runtime bolted on the side (Lua, or Caddy's plugins). Behavior ends up split across two layers: directives that quietly grow their own control flow, plus scripts that run somewhere in the request lifecycle you have to keep in your head.
I think the bet is misplaced - people prefer configuration over code and long have. The built-ins meet enough peoples needs entirely and they don't need to write C code.
I'd wager that in the next 96 hours, with a LLM, someone could create a translator that would 'pack' a nginx or caddy configuration file into the relevant code that zeroserve could use. Or even more simply, just pickup all the Ingress manifests in a kubernetes cluster and rebuild the pack. The point being, the interface between the tool and the configuration is just another API, system operators are already describing the state of the system at higher level constructs, and the specific bytes that make up the configuration are an artifact of that.
I suspect it's worth exploring if that will change as AI is allowing more and more "human words -> machine effect", which may be more ergonomic for the AIs. It could take a long time for that kind of shift to become clearly a good idea since the AI can make either work.
Why are you so eager to credit LLMS.. Just because they wrote the article with the help of an LLM doesnt mean they're out here having LLMs do the experiments for them.
I think I'd feel more comfortable if I could drop an .rs file into the eBPF dir instead of a .c one. It's already a Rust project! :)
And for some reason I was expecting this to be a kernel-accelerated webserver - if that could be done safely using eBPF that would be amazing!
Also, single-threaded? Forking and sharing an incoming connection queue is basically trivial on Linux, that should be literally just a few lines, even with Rust. Use SO_REUSEPORT and the kernel will do the rest.
FWIW, if you're going to push for io_uring, you should also be pushing kTLS IMO, you'll drastically simplify your design if you can avoid pumping userspace SSL after the handshake.
Very cool! would be interesting to see about combining this with other bpf program types like xdp progs, or socket map attached programs to integrate L7 http features downward.
I did last week (converted ghost to static) and was half wondering if some self contained binary wouldn't be faster so I feel like this was made for me, but I accept I'm not the typical user
These projects are fascinating, and I referenced them in a nearby comment about static hosting from archives. I need to try the latest versions to see how they work at higher scale (more data in the archive).
Depends on the domain. There's a bunch of sciences using large datasets served up efficiently using static file formats, e.g., https://zarr.dev/ and https://parquet.apache.org/
Looks good, nice features. But somehow the spark does not ignite on my side because it feels too artificial. I don't know if the metrics are faked, if the convenience functions actually work, if there is any proper hardening.
I can accept if stuff is vibe coded and has autogenerated README. But even the announcement blogpost is AI-generated, and I personally have zero data points to see if your understanding of software quality is the same as mine.
It's a weird world, if this would've been announced without any AI disclaimers some years earlier I would've eaten it up without a doubt. But right now if I see a fancy README with several good-looking command line parameters I immediately wonder if the README is hallucinated and the command line parameters actually exist.
Hi, author here - a few critical pieces of this, like async-ebpf, were written long before those coding agents were released. I use AI assistance a lot when creating zeroserve itself, but I manually check AI output and take responsibility for it :)
I'm of the school of thought that if a practicing/retired software engineer (i.e. someone I reasonably believe has experience writing software for "production") wrote it, I've got to show it's trash, rather than assume it's trash. "Innocent until proven guilty" and all that. But I'm in the rather luxurious position of mostly using open source, rather than maintaining it, so I understand that others come down differently on this topic.
FWIW, I like the writeup and concept behind this. Very close to some passions of mine (like serving a website from a single-file archive).
if the point is to avoid the lua-issue on nginx, how do you expect people will implement things like geoip, request content match post ssl termination, etc?
> It's a weird world, if this would've been announced without any AI disclaimers some years earlier I would've eaten it up without a doubt. But right now if I see a fancy README with several good-looking command line parameters I immediately wonder if the README is hallucinated and the command line parameters actually exist.
Yeah, that is unfortunate. Recently there was this ffmpeg-wasm project. I tested it. It worked. But it was vibe-coded AI. I can't stand AI. Even if things work.
I decided to stay in the oldschool era as much as possible. Clever people publish software. Clever people maintain software. They don't need AI. That's my niche.
We may die out but I still prefer that. (Oh, and only if these clever people write documentation. Many clever people hate writing documentation. I decided a long time ago that if software comes without documentation, it is not worth my time, no matter how great that documentation is. This refers mostly to on-the-application side; I only rarely looked at the Linux documentation, but others stated that it is not too terrible either, so who knows.)
Small static file (174 B) - the bread and butter of static sites:
server req/s p99
zeroserve 36,681 5.4 ms
nginx 31,226 7.8 ms
Caddy 12,830 22 ms
zeroserve serves small files about 17% faster than nginx on a single core, with a tighter tail. HTML pages, small JSON, CSS - this is the case zeroserve is tuned for.
Large static file (100 KB):
server req/s throughput p99
zeroserve 8,000 782 MB/s 22 ms
nginx 7,600 773 MB/s 28 ms
Caddy 6,084 590 MB/s 44 ms
I'd go with a more storied project that's been audited, battle tested, hardened etc than this upstart. There's not enough improvement to justify the risk.
The problem with pasting LLM output is that no human with sound mind and body would waste their finite time on this Earth informing you that small static files are "the bread and butter of static sites".
It would be interesting to extend it to zip, which is what redbean/greenbean use to serve static assets.
Back in school, I worked on a project called Velox, with a partner - the idea was to take a bz2-compressed dump of the giant XML export of wikipedia, and write a program to serve that copy of wikipedia from disk (this was in 2008-2010? in my master's program, so before Kiwix and the amazing zim dumps they produce). My partner worked on the UI and indexing, and I was focusing on how to parse the bz2 compression format to locate article boundaries in the (giant) XML dump that Wikipedia provides. I ended up putting a lot of time into it because it was a bunch of fun.
Writing this just sent me back to the presentation I made. The slide I wrote back then said:
> Significant original work went into creation of archive access. The Apache BZip2 library that is part of Ant was used as a basis for archive access.
> Modified to support random access to a given byte/bit offset pair within the compressed data stream (BZip2 is not a byte-aligned format)
> Extended to index all BZip2 block positions, allowing Java-based pseudo-random access to BZip2 compressed data
> Extended to map article IDs to block numbers for constant-time article retrieval, even in BZip2 archives exceeding 5GB in size
> Current article retrieval times are ~2 seconds.
This is back when the archive was ~7GB IIRC. My Kiwix dumps today are ~120GB, but that includes images.
This is the link to the presentation in Google Slides that we wrote back in 2008 or so. The version history shows 2013, but I think some kind of import/conversion happened around that time.
Well, according to first paragraph of the section titled "One tarball, served in place":
The whole site is a single tar file. zeroserve indexes it on load - building a path -> byte-range map - and then serves files by issuing byte-range reads against the tarball itself. Nothing is ever unpacked to disk. The site lives entirely in that one file, so there's no document root for a stray location rule to expose, and a deploy is a single atomic file swap.
OTOH, that could be an LLM justification, since the copy is littered with -isms like "the right shape" or "the surface is broad".
The real question is developer commitment and community - the Caddy and Nginx people have worked constantly on supporting their products. It’s going to take a lot of focus and attention.
> All you did was ask the slop generator to make it.
Is this founded? I know it's popular to trash every new project that's built with AI in the pipeline somewhere, but there's a big difference between a project built by someone with years of experience writing software and some vibing a repo with Claude. Isn't it worth distinguishing between the two?
In the case of this project: a non-technical person couldn't even conceive of this product. So I think some kind of evidence should be presented before we say things like "Is it a logical design? Who knows."
The death of the techempower web server benchmarks means new ones like this one no longer have the chance to prove themselves.
Edit: it seems I'm just falling behind and the new hotness is https://www.http-arena.com/leaderboard/. Good luck!
I love seeing stuff like this that would probably not exist if not for LLMs making exploring these kinds of ideas relatively cheap and quick to do.
My takeaway from this though is that nginx is pretty impressive on its own. Also this stuck out:
It's meant to be an alternative to nginx and Caddy, and the design bet is about configuration. Those servers give you a declarative config language - location blocks, rewrite rules, map directives, try_files - and then, once the declarative language hits its limits, an optional scripting runtime bolted on the side (Lua, or Caddy's plugins). Behavior ends up split across two layers: directives that quietly grow their own control flow, plus scripts that run somewhere in the request lifecycle you have to keep in your head.
I think the bet is misplaced - people prefer configuration over code and long have. The built-ins meet enough peoples needs entirely and they don't need to write C code.
I'd wager that in the next 96 hours, with a LLM, someone could create a translator that would 'pack' a nginx or caddy configuration file into the relevant code that zeroserve could use. Or even more simply, just pickup all the Ingress manifests in a kubernetes cluster and rebuild the pack. The point being, the interface between the tool and the configuration is just another API, system operators are already describing the state of the system at higher level constructs, and the specific bytes that make up the configuration are an artifact of that.
What about abstracting the complexity and achieving "config file" configuration with macros?
I suspect it's worth exploring if that will change as AI is allowing more and more "human words -> machine effect", which may be more ergonomic for the AIs. It could take a long time for that kind of shift to become clearly a good idea since the AI can make either work.
Why are you so eager to credit LLMS.. Just because they wrote the article with the help of an LLM doesnt mean they're out here having LLMs do the experiments for them.
The commit history shows plenty of coding agent involvement: https://github.com/losfair/zeroserve/commits/main/
because it makes it easy to create/reverse engineer projects/explore on a whim without putting up a lot of time
LLM enables a lot of good output if you know what you are doing
Bayesian inference suggests otherwise.
I like the idea.
I think I'd feel more comfortable if I could drop an .rs file into the eBPF dir instead of a .c one. It's already a Rust project! :)
And for some reason I was expecting this to be a kernel-accelerated webserver - if that could be done safely using eBPF that would be amazing!
Also, single-threaded? Forking and sharing an incoming connection queue is basically trivial on Linux, that should be literally just a few lines, even with Rust. Use SO_REUSEPORT and the kernel will do the rest.
FWIW, if you're going to push for io_uring, you should also be pushing kTLS IMO, you'll drastically simplify your design if you can avoid pumping userspace SSL after the handshake.
Hi, thanks!
Will implement forking + SO_REUSEPORT. I've been using nftables for things like this so haven't needed it for myself yet :)
Very cool! would be interesting to see about combining this with other bpf program types like xdp progs, or socket map attached programs to integrate L7 http features downward.
Cool idea, but I don't think you should focus on static files. People rarely spin up a server for that these days.
I did last week (converted ghost to static) and was half wondering if some self contained binary wouldn't be faster so I feel like this was made for me, but I accept I'm not the typical user
You just helped to dredge up a memory, which brought me back to this fascinating project:
https://redbean.dev
If this piques your interest, make sure to check out the portable C library used to create it, which is also fascinating:
https://github.com/jart/cosmopolitan
These projects are fascinating, and I referenced them in a nearby comment about static hosting from archives. I need to try the latest versions to see how they work at higher scale (more data in the archive).
Why not let Clouflare/GitHub/etc do it for you? Free, and you don't have to worry about security and availability.
Depends on the domain. There's a bunch of sciences using large datasets served up efficiently using static file formats, e.g., https://zarr.dev/ and https://parquet.apache.org/
Admit I'm not familiar with that domain. But don't people use managed services even there? In my job we host parquet files on S3.
I did it yesterday.
Looks good, nice features. But somehow the spark does not ignite on my side because it feels too artificial. I don't know if the metrics are faked, if the convenience functions actually work, if there is any proper hardening.
I can accept if stuff is vibe coded and has autogenerated README. But even the announcement blogpost is AI-generated, and I personally have zero data points to see if your understanding of software quality is the same as mine.
It's a weird world, if this would've been announced without any AI disclaimers some years earlier I would've eaten it up without a doubt. But right now if I see a fancy README with several good-looking command line parameters I immediately wonder if the README is hallucinated and the command line parameters actually exist.
Hi, author here - a few critical pieces of this, like async-ebpf, were written long before those coding agents were released. I use AI assistance a lot when creating zeroserve itself, but I manually check AI output and take responsibility for it :)
I'm of the school of thought that if a practicing/retired software engineer (i.e. someone I reasonably believe has experience writing software for "production") wrote it, I've got to show it's trash, rather than assume it's trash. "Innocent until proven guilty" and all that. But I'm in the rather luxurious position of mostly using open source, rather than maintaining it, so I understand that others come down differently on this topic.
FWIW, I like the writeup and concept behind this. Very close to some passions of mine (like serving a website from a single-file archive).
Happy to hear, I hope the tool can prove itself to a wider audience then.
if the point is to avoid the lua-issue on nginx, how do you expect people will implement things like geoip, request content match post ssl termination, etc?
> It's a weird world, if this would've been announced without any AI disclaimers some years earlier I would've eaten it up without a doubt. But right now if I see a fancy README with several good-looking command line parameters I immediately wonder if the README is hallucinated and the command line parameters actually exist.
Yeah, that is unfortunate. Recently there was this ffmpeg-wasm project. I tested it. It worked. But it was vibe-coded AI. I can't stand AI. Even if things work.
I decided to stay in the oldschool era as much as possible. Clever people publish software. Clever people maintain software. They don't need AI. That's my niche.
We may die out but I still prefer that. (Oh, and only if these clever people write documentation. Many clever people hate writing documentation. I decided a long time ago that if software comes without documentation, it is not worth my time, no matter how great that documentation is. This refers mostly to on-the-application side; I only rarely looked at the Linux documentation, but others stated that it is not too terrible either, so who knows.)
Given the benchmarks:
Small static file (174 B) - the bread and butter of static sites:
server req/s p99
zeroserve 36,681 5.4 ms
nginx 31,226 7.8 ms
Caddy 12,830 22 ms
zeroserve serves small files about 17% faster than nginx on a single core, with a tighter tail. HTML pages, small JSON, CSS - this is the case zeroserve is tuned for.
Large static file (100 KB):
server req/s throughput p99
zeroserve 8,000 782 MB/s 22 ms
nginx 7,600 773 MB/s 28 ms
Caddy 6,084 590 MB/s 44 ms
I'd go with a more storied project that's been audited, battle tested, hardened etc than this upstart. There's not enough improvement to justify the risk.
The problem with pasting LLM output is that no human with sound mind and body would waste their finite time on this Earth informing you that small static files are "the bread and butter of static sites".
I'm convinced that LLMs somehow settled on the middle manager as the exemplar of human cognition that it tries its best to emulate.
I could totally see "Small static files are the bread and butter of static sites" appearing in some pointless deck on a Zoom call.
Why a tarball?
It's a simple format easily suitable for accessing resources by byte ranges, that everyone has tooling for, and which _doesn't_ compress things.
It would be interesting to extend it to zip, which is what redbean/greenbean use to serve static assets.
Back in school, I worked on a project called Velox, with a partner - the idea was to take a bz2-compressed dump of the giant XML export of wikipedia, and write a program to serve that copy of wikipedia from disk (this was in 2008-2010? in my master's program, so before Kiwix and the amazing zim dumps they produce). My partner worked on the UI and indexing, and I was focusing on how to parse the bz2 compression format to locate article boundaries in the (giant) XML dump that Wikipedia provides. I ended up putting a lot of time into it because it was a bunch of fun.
Writing this just sent me back to the presentation I made. The slide I wrote back then said:
> Significant original work went into creation of archive access. The Apache BZip2 library that is part of Ant was used as a basis for archive access.
> Modified to support random access to a given byte/bit offset pair within the compressed data stream (BZip2 is not a byte-aligned format) > Extended to index all BZip2 block positions, allowing Java-based pseudo-random access to BZip2 compressed data > Extended to map article IDs to block numbers for constant-time article retrieval, even in BZip2 archives exceeding 5GB in size
> Current article retrieval times are ~2 seconds.
This is back when the archive was ~7GB IIRC. My Kiwix dumps today are ~120GB, but that includes images.
This is the link to the presentation in Google Slides that we wrote back in 2008 or so. The version history shows 2013, but I think some kind of import/conversion happened around that time.
https://docs.google.com/presentation/d/e/2PACX-1vTfrxEqvHbd0...
Well, according to first paragraph of the section titled "One tarball, served in place":
The whole site is a single tar file. zeroserve indexes it on load - building a path -> byte-range map - and then serves files by issuing byte-range reads against the tarball itself. Nothing is ever unpacked to disk. The site lives entirely in that one file, so there's no document root for a stray location rule to expose, and a deploy is a single atomic file swap.
OTOH, that could be an LLM justification, since the copy is littered with -isms like "the right shape" or "the surface is broad".
Thanks, I missed that during my read of it
It’s an interesting new concept I like it.
The real question is developer commitment and community - the Caddy and Nginx people have worked constantly on supporting their products. It’s going to take a lot of focus and attention.
this looks amazing
[flagged]
You should post the results of your source code audit.
> All you did was ask the slop generator to make it.
Is this founded? I know it's popular to trash every new project that's built with AI in the pipeline somewhere, but there's a big difference between a project built by someone with years of experience writing software and some vibing a repo with Claude. Isn't it worth distinguishing between the two?
In the case of this project: a non-technical person couldn't even conceive of this product. So I think some kind of evidence should be presented before we say things like "Is it a logical design? Who knows."