People have been using pcntl_fork to speed up PHP code since the 90s. It seems good on paper, can produce an "impressive" demo and falls apart quickly on real world projects and use cases. This is nothing more than LLM slop/psychosis.
The general idea makes sense. PHP, particularly modern Java-style PHP, is notorious for loading sometimes thousands of files for every single request. I think there's opcache enabled by default, so the load on the filesystem is reduced by quite a bit, but even opcache still needs to do some sort of parsing for each request that comes in.
Indeed, it's on by default in 8.5 and is no longer considered a non-optional extension (it's compiled in) - you can turn it off at runtime but it will be bundled, largely so they could have it on my default[1][2] and it simplifies integration of something that was universally integrated de facto before that.
It was probably due/overdue but I actually like that they are conservative with which extensions become part of the "core" and in reality almost every packaging of PHP has included it back to 5.5 and having worked on large production systems written in PHP I can't think of a single one that didn't use it it in production.
There is also FrankenPHP which negates much of that overhead (in worker mode) from building the world to throw it away.
While introducing other issues you do have to be careful of that would be less of an issue in "classic" mode (literally what FrankenPHP calls it).
The build and throw away the world solution does mitigate/obviate some issues but has it's own issues (memory leaks and the like become more of an issue in long lived worker processes), as always the final answer is "it depends".
There isn't a tonne of good info out of what you gain in the switch from nginx/php-fpm to FrankenPHP though it does seem like you do gain in some ways (but it's very application dependent).
You don't need your database on a dedicated server. Not sure why the comparison table claims "$30 + DB server" when your web server is a perfectly good server already.
Presumably to remove it from the need to calculate. Everyone is using a different approach and one DB may require a lot more than the 30$ server can provide.
The claim is presumably that a web server running SQLite is that much better than a web server running Postgres or Redis. So they end up including it in the calculation anyway but only on one side... so it's just simply not even apples to apples.
I for one do not get the hate, nobody needs to use/or-waste-time-commenting on this project if they have no need/want. However, I do agree that it needs some more contributors on-board alongside with a couple months more support and bug squashing. And yes: better rewrite that website without the clanger involved in the wording! ;)
On a positive note: the performance is genuinely impressive.. what does it take to unlock this performance if I'm NOT running one of the big "usual" frameworks? The devil is always in the detail, best get some practical advice up on exactly what needs to happen with a medium piece of custom code to get it to said max performance.
I think we are all afraid that an LLM could generate a ridiculously fast webserver and be production viable. But I'm still glad we are taking this through the ringer.
I was kind of wondering too. I assume this only works if you load a lot of classes before touching request specific variables. That might work for big framework, but you probably have to rewrite a lot if not using one.
From a pragmatic perspective, its unclear to me how much this actually speeds things up. I would assume if using php you probably have complex logic and start up time is not dominating the runtime.
I'm confused why the copy you generated repeatedly boasts "no nginx, no Redis" as if those things don't exist independently of application servers for a reason, and as if building something that doesn't need these things (when deployed as a single instance) is particularly notable.
"no Docker" - so what? Docker is a tool. Running containers on Linux has extremely low overhead. No idea what this is about.
You didn't spend the effort to decorate the slop with your rationale for these choices, so I see no evidence that you built any of this.
One of the biggest shortcomings I often see with LLM-produced content, following a planning conversation, is that they misunderstand the breadth of the domain for which a detail is relevant or interesting.
So if it fixes a bug where function was accidentally deleting a file, it will update the function's documentation with "does not delete any important files", as if that's a key feature and not just one of a million things it should not do.
The same pattern repeats when an LLM is left to generate its own copy after a brainstorming and architecting session. It has no concept of what's important to the end-user.
The only LLM that does this wel (in my opinion) is gemini. if i ask claude to fix comments, it very clearly fails in doing so. Gemini actually seems to understand the scope of a task
It's generally easier to have language self-host itself because at scale, you will be running multiple instances of the application so instead of Ingress Controller -> Go self-hosted/Python Uvicorn, it becomes Ingress COntroller -> Apache -> PHP and if developer modifies the Apache in ways you are not expecting, stuff can break.
I mean at scale you probably have TLS termination -> caching layer -> webserver -> fpm -> php. All probably written in different languages.
I find this argument kind of unconvincing. At scale you want your components to be rock solid. This project is an interesting experiment, but i would never use it at scale until it matures a lot more.
I agree. Nginx, Caddy, Apache, etc have found and fixed a million edge cases you'd never think of just from reading and implementing the spec. I guarantee there's code in each of these along the lines of:
# Send an extra 0x20 space character after this header's value, because
# otherwise Chrome on Android 15 shifts into compatibility mode and
# it takes 37x longer to render the page, which everyone will blame on
# "the slow webserver". The spec doesn't say we *can't* add this, and
# tells clients to ignore trailing space, and we've tested this with 483
# other clients to demonstrate that it doesn't cause problems. ¯\_(ツ)_/¯ .
Now, as a very cool hack, or for easier local development with one less dependency, or just to scratch a personal itch and see if it can even be done, right on! That's clever and I'm glad they did it!
BTW, in your showdead (for wholly unclear reasons) comment about it using copy-on-write RAM semantics, that's just the Unix process model. The OS does all that courtesy of fork() and the server gets it for free.
You can see the slop through the headline. No need to click. Posts like this are turning the internet into a trash pile. Nobody needs this slop, abandonware.
Heres the template:
"I" <--- prompted some LLM
"built" <--- to generate code for a
"fastest" <-- insert specious claim here
"PHP server" <-- thing
"And I prompted the LLM to build a website" <---- link
Im sorry folks but this just doesnt cut it anymore. I dont think it ever did.
People have been using pcntl_fork to speed up PHP code since the 90s. It seems good on paper, can produce an "impressive" demo and falls apart quickly on real world projects and use cases. This is nothing more than LLM slop/psychosis.
What do you mean LLM slop? OP says they built it themselves.
Yeah, and surely they also wrote the 131KB unreadable-wall-of-text README themselves.
That's like 1/9th Moby Dick, to put into perspective.
You should look at Workerman
Workerman does something else.
I love the idea of being able to remove the webserver from my stack, and be able to give that job to php too. Less deps!
s/you built/you spent LLM tokens to build/g
The general idea makes sense. PHP, particularly modern Java-style PHP, is notorious for loading sometimes thousands of files for every single request. I think there's opcache enabled by default, so the load on the filesystem is reduced by quite a bit, but even opcache still needs to do some sort of parsing for each request that comes in.
> I think there's opcache enabled by default.
Indeed, it's on by default in 8.5 and is no longer considered a non-optional extension (it's compiled in) - you can turn it off at runtime but it will be bundled, largely so they could have it on my default[1][2] and it simplifies integration of something that was universally integrated de facto before that.
It was probably due/overdue but I actually like that they are conservative with which extensions become part of the "core" and in reality almost every packaging of PHP has included it back to 5.5 and having worked on large production systems written in PHP I can't think of a single one that didn't use it it in production.
There is also FrankenPHP which negates much of that overhead (in worker mode) from building the world to throw it away.
While introducing other issues you do have to be careful of that would be less of an issue in "classic" mode (literally what FrankenPHP calls it).
The build and throw away the world solution does mitigate/obviate some issues but has it's own issues (memory leaks and the like become more of an issue in long lived worker processes), as always the final answer is "it depends".
There isn't a tonne of good info out of what you gain in the switch from nginx/php-fpm to FrankenPHP though it does seem like you do gain in some ways (but it's very application dependent).
[1] https://wiki.php.net/rfc/make_opcache_required
[2] https://www.php.net/manual/en/opcache.configuration.php
You don't need your database on a dedicated server. Not sure why the comparison table claims "$30 + DB server" when your web server is a perfectly good server already.
Why do you need a dedicated server at all? A $7 droplet/vps would handle the full stack
Presumably to remove it from the need to calculate. Everyone is using a different approach and one DB may require a lot more than the 30$ server can provide.
Now everyone can understand the math.
The claim is presumably that a web server running SQLite is that much better than a web server running Postgres or Redis. So they end up including it in the calculation anyway but only on one side... so it's just simply not even apples to apples.
I for one do not get the hate, nobody needs to use/or-waste-time-commenting on this project if they have no need/want. However, I do agree that it needs some more contributors on-board alongside with a couple months more support and bug squashing. And yes: better rewrite that website without the clanger involved in the wording! ;)
On a positive note: the performance is genuinely impressive.. what does it take to unlock this performance if I'm NOT running one of the big "usual" frameworks? The devil is always in the detail, best get some practical advice up on exactly what needs to happen with a medium piece of custom code to get it to said max performance.
I think we are all afraid that an LLM could generate a ridiculously fast webserver and be production viable. But I'm still glad we are taking this through the ringer.
I was kind of wondering too. I assume this only works if you load a lot of classes before touching request specific variables. That might work for big framework, but you probably have to rewrite a lot if not using one.
From a pragmatic perspective, its unclear to me how much this actually speeds things up. I would assume if using php you probably have complex logic and start up time is not dominating the runtime.
I'm confused why the copy you generated repeatedly boasts "no nginx, no Redis" as if those things don't exist independently of application servers for a reason, and as if building something that doesn't need these things (when deployed as a single instance) is particularly notable.
"no Docker" - so what? Docker is a tool. Running containers on Linux has extremely low overhead. No idea what this is about.
You didn't spend the effort to decorate the slop with your rationale for these choices, so I see no evidence that you built any of this.
One of the biggest shortcomings I often see with LLM-produced content, following a planning conversation, is that they misunderstand the breadth of the domain for which a detail is relevant or interesting.
So if it fixes a bug where function was accidentally deleting a file, it will update the function's documentation with "does not delete any important files", as if that's a key feature and not just one of a million things it should not do.
The same pattern repeats when an LLM is left to generate its own copy after a brainstorming and architecting session. It has no concept of what's important to the end-user.
The only LLM that does this wel (in my opinion) is gemini. if i ask claude to fix comments, it very clearly fails in doing so. Gemini actually seems to understand the scope of a task
In the specific context though, it is relevant. So the aside is more closely related to the behavior being outlined.
Really wish I could use Gemini without linking it to my personal Google account.
In the context of PHP it is notable to be able to run a production grade web server without nginx/apache/docker in front of it.
PHP doesn't have a production grade built-in server like .NET's Kestrel.
PHP is stable enough to not need Docker.
As in there’s very little ways to fuck up a PHP setup.
You just turn it on and it runs.
then find out you need to install and configure a series of php extensions
I haven't used PHP as a daily workhorse in a decade, so I guess I don't understand why I would want a web server in pure PHP? Nginx exists.
It's generally easier to have language self-host itself because at scale, you will be running multiple instances of the application so instead of Ingress Controller -> Go self-hosted/Python Uvicorn, it becomes Ingress COntroller -> Apache -> PHP and if developer modifies the Apache in ways you are not expecting, stuff can break.
I mean at scale you probably have TLS termination -> caching layer -> webserver -> fpm -> php. All probably written in different languages.
I find this argument kind of unconvincing. At scale you want your components to be rock solid. This project is an interesting experiment, but i would never use it at scale until it matures a lot more.
I agree. Nginx, Caddy, Apache, etc have found and fixed a million edge cases you'd never think of just from reading and implementing the spec. I guarantee there's code in each of these along the lines of:
Now, as a very cool hack, or for easier local development with one less dependency, or just to scratch a personal itch and see if it can even be done, right on! That's clever and I'm glad they did it!BTW, in your showdead (for wholly unclear reasons) comment about it using copy-on-write RAM semantics, that's just the Unix process model. The OS does all that courtesy of fork() and the server gets it for free.
You can see the slop through the headline. No need to click. Posts like this are turning the internet into a trash pile. Nobody needs this slop, abandonware.
Heres the template:
"I" <--- prompted some LLM
"built" <--- to generate code for a
"fastest" <-- insert specious claim here
"PHP server" <-- thing
"And I prompted the LLM to build a website" <---- link
Im sorry folks but this just doesnt cut it anymore. I dont think it ever did.
I think that is far to harsh actually.
It has some interesting ideas and architecture.
I feel like HN has had posts like this for 15 years. This one is just an opportunity to discuss where php falls with llms in the mix.
I saw someone promote their new simple framework today. It was all TODOs.
Just because LLMs can write copy for webpages doesn't mean they should. A landing page should not warrant a tl;dr.
https://news.ycombinator.com/item?id=49665895