Why I forked httpx
tildeweb.nl
tildeweb.nl
I feel like it's counterproductive in situations like this to mention forking. It will come across like a threat, when there isn't really anything intrinsically aggressive about it. So just do it; and when you have a decent amount of separate development, you can decide whether to make PRs back, advertise your fork, etc.
I am trying to resolve what you've seen. For years of hard work.
I think that it owes its success to be first "port" of python requests to support async, that was a strong need.
But otherwise it is bad: API is not that great, performance is not that great, tweaking is not that great, and the maintainer mindset is not that great also. For the last point, few points were referenced in the article, but it can easily put your production project to suddenly break in a bad way without valid reason.
Without being perfect, I would advise everyone to switch to Aiohttp.
fwiw, HTTP/2 is twelve years old, just saying.
Basically treating HTTP requests as an orthogonal, or cross-cutting concern.
It is sometimes hard to tell if these upstream packages are stable or abandoned.
I should probably document my methodology so it can help others or at least have the chance to find out what mistakes or limitations they might have.
more seriously, all that is needed is our collective effort. I've done my part by scarifying a lot of personal time for it.
I've started seeing these emoji-prefixed commits lately now too, peculiar
J'aime bien, j'aime bien
(Also the sponsorship subscription thing in the readme gives me vague rugpull vibes. Maybe I’ve just been burned too much—I don’t mean to discourage selling support in general.)
It's been a pleasure to use, has a httpx compatibility layer for gradually migrating to its API, and it's a lot more performant (right now, I think it's the most performant Python http client out there: https://github.com/MarkusSintonen/pyreqwest/blob/main/docs/b...)
I still think that hijacking the mkdocs package was the wrong way to go though.
The foss landscape has become way too much fork-phobic.
Just fork mkdocs and go over your merry way.
Pretty sure those are completely standard for major changes in maintainers/hostile forks/acknowledging major contributors. I've seen a lot of abandoned MIT/BSD projects add a new line for forks/maintainers being active again in order to acknowledge that the project is currently being headed by someone else.
From my "I am not a lawyer" view, Kludex is basically correct, although I suppose to do it "properly", he might need to just duplicate the license text in order to make it clear both contributors licensed under BSD 3-clause. Probably unnecessary though, given it's not a license switch (you see that style more for ie. switching from MIT to BSD or from MIT/BSD to GPL, since that's a more substantial change); the intent of the license remains the same regardless and it's hard to imagine anyone would get confused.
I suspect (given the hammering on it in responses), that Kludex asking ChatGPT if it was correct is what actually pissed off the original developer, rather than the addition of Kludex to the list in and of itself.
The original author said they were “the license holder”, specifically with a “the”, in discussions around both Starlette and MkDocs, which yes, just isn’t true even after rounding the phrase to the nearest meaningful, “the copyright holder”. This appears to be an honest misconception of theirs, so, not the end of the world, except they seem to be failing at communication hard enough to not realize they might be wrong to begin with.
Note though that with respect to Starlette this ended up being essentially a (successful and by all appearances not intentionally hostile?) project takeover, so the emotional weight of the drama should be measured with respect to that, not just an additional copyright line.
Mia Kimberly Christie seems like dangerous person.
[1] https://github.com/orgs/encode/discussions/11#discussioncomm...
I don’t think that’s the case. It’s more of a marketing/market incentive. It’s great pr to be associated with the most famous project, way less so to be associated with a fork, at least until the fork becomes widespread and well recognised.
GitHub does make it fairly easy to fork a project, I wouldn’t blame the situation on github.
yeah i'm gonna call BS on this. this kind of drama already existed when communities were "decentralised" and each one had its own forum, mailing list or whatever.
the core of the issue here is about wanting to be the owner of a repository.
so people should just not bother with being owner of a specific repository, but just fork it and move on. and github supports forking sufficiently well for this purpose.
1. Issues, PRs, discussions etc don't move over,
2. Discovery. When GitHub was in its infancy they made a big deal out of the "network" feature. Almost 20 years later it has languished, virtually unchanged in that entire time. When I think of decentralisation I think Bitcoin. We need a better way to "discover" the consensus fork. I feel like GitHub could do a lot to help, but they don't care.
- lovelydinosaur,
- Mia Kimberly Christie
- Kim (Christie)
- Tom Christie
>They also point out that not opening up the source code goes against the principles of Open Source software development
I will never stop being amused when people have feelings like this and also choose licenses like BSD (this project). If you wanted a culture that discouraged those behaviors, why would you choose a license that explicitly allows them? Whether you can enforce it or not, the license is basically a type of CoC that states the type of community you want to have.
This usually doesn't work, and in the end all they can do is complain about behaviours that their license choice explicitly allowed.
On the other hand, the comments the MkDocs author is making about perceived gender grievances feel so unhinged that I wouldn't be touching anything made by them with a barge pole.
Oleh was basically the sole maintainer for many years, and the development basically stopped when he left.
This seems like a pretty good reason to fork to me.
> Sending HTTP requests is a basic capability in the modern world, the standard library should include a friendly, fully-featured, battle-tested, async-ready client. But not in Python,
Or Javascript (well node), or golang (http/net is _worse_ than urllib IMO), Rust , Java (UrlRequest is the same as python's), even dotnet's HttpClient is... fine.
Honestly the thing that consistently surprises me is that requests hasn't been standardised and brought into the standard library
HttpClient client = HttpClient.newBuilder()
.version(Version.HTTP_1_1)
.followRedirects(Redirect.NORMAL)
.connectTimeout(Duration.ofSeconds(20))
.proxy(ProxySelector.of(
new InetSocketAddress("proxy.example.com", 80)
))
.authenticator(Authenticator.getDefault())
.build();
HttpResponse<String> response = client.send(request, BodyHandlers.ofString());
System.out.println(response.statusCode());
System.out.println(response.body());
using HttpClient client = new();
HttpResponseMessage response = await client.GetAsync("https://...");
if (response.StatusCode is HttpStatusCode.OK)
{
string s = await response.Content.ReadAsStringAsync();
// ...
}"Common IHttpClientFactory usage issues"
https://learn.microsoft.com/en-us/dotnet/core/extensions/htt...
"Guidelines for using HttpClient"
https://learn.microsoft.com/en-us/dotnet/fundamentals/networ...
And this doesn't account for all gotchas as per .NET version, than only us old timers remember to cross check.
> What's the matter with this?
To me what makes this very "Java" is the arguments being passed, and all the OOP stuff that isn't providing any benefit and isn't really modeling real-world-ish objects (which IMHO is where OOP shines). .version(Version.HTTP_1_1) and .followRedirects(Redirect.NORMAL) I can sort of accept, but it requires knowing what class and value to pass, which is lookups/documentation reference. These are spread out over a bunch of classes. But we start getting so "Java" with the next ones. .connectTimeout(Duration.ofSeconds(20)) (why can't I just pass 20 or 20_000 or something? Do we really need another class and method here?) .proxy(ProxySelector.of(new InetSocketAddress("proxy.example.com", 80))), geez that's complex. .authenticator(Authenticator.getDefault()), why not just pass bearer token or something? Now I have to look up this Authenticator class, initialize it, figure out where it's getting the credentials, how it's inserting them, how I put the credentials in the right place, etc. The important details are hidden/obscured behind needless abstraction layers IMHO.
I think Java is a good language, but most modern Java patterns can get ludicrous with the abstractions. When I was writing lots of Java, I was constantly setting up an ncat listener to hit so I could see what it's actually writing, and then have to hunt down where a certain thing is being done and figuring out the right way to get it to behave correctly. Contrast with a typical Typescript HTTP request and you can mostly tell just from reading the snippet what the actual HTTP request is going to look like.
If you've ever dealt with time, you'll be grateful it's a duration and not some random int.
Unless you use a text editor without any coding capabilities, your IDE should show you which values you can pass. The alternative is to have more methods, I guess?
> why can't I just pass 20 or 20_000 or something
20 what? Milliseconds? Seconds? Minutes? While I wouldn't write the full Duration.ofSeconds(20) (you can save the "Duration."), I don't understand how one could prefer a version that makes you guess the unit.
> proxy(ProxySelector.of(new InetSocketAddress("proxy.example.com", 80))), geez that's complex
Yes it is, can't add anything here. There's a tradeoff between "do the simple thing" and "make all things possible", and Java chooses the second here.
> .authenticator(Authenticator.getDefault()), why not just pass bearer token or something?
Because this Authenticator is meant for prompting a user interactively. I concur that this is very confusing, but if you want a Bearer token, just set the header.
> Unless you use a text editor without any coding capabilities, your IDE should show you which values you can pass. The alternative is to have more methods, I guess?
Fair enough, as much as I don't like it, in Java world it's safe to assume everyone is using an IDE. And when your language is (essentially) dependent on an IDE, this becomes a non-issue (actually I might argue it's even a nice feature since it's very type safe).
> 20 what? Milliseconds? Seconds? Minutes? While I wouldn't write the full Duration.ofSeconds(20) (you can save the "Duration."), I don't understand how one could prefer a version that makes you guess the unit.
I would assume milliseconds and would probably have it in the method name, like timeoutMs(...) or something. I will say it's very readable, but if I was writing it I'd find it annoying. But optimizing for readability is a reasonable decision, especially since 80% of coding is reading rather than writing (on average).
I feel like you answered yourself. Java makes you do this by not supporting proper keyword arguments.
HttpClient client = HttpClient.newBuilder()
.followRedirects(Redirect.NORMAL)
.connectTimeout(Duration.ofSeconds(20))
.build();It does enable code that "hides" implementation very well, like the quoted examples authentication API lets you authenticate in any way you can imagine, as in literally any way imaginable.
Its incredibly flexible. Want to only be able to send the request out after you've touched a file, send of a Message through a message broker and then maybe flex by waiting for the response of that async communication and use that as a custom attribute in the payload, additionally to a dynamically negotiated header to be set according to the response of a DNS query? yeah, we can do that! and the caller doesnt have to know any of that... at least as long as it works as intended
Same with the Proxy layer, the client is _entirely_ extensible, it is what Inversion of Control enables.
It just comes with the unfortunate side-effect of forcing the dev to be extremely fluent in enterprisey patterns. I dont mind it anymore, myself. the other day ive even implemented a custom "dependency injection" inspired system for data in a very dynamic application at my dayjob. I did that so the caller wont even need to know what data he needs! it just get automatically resolved through the abstraction. But i strongly suspect if a jr develeoper which hasnt gotten used to the java ecosystem will come across it, he'll be completely out of his depth how the grander system works - even though a dev thats used to it will likely understand the system within a couple of moments.
Like everything in software, everything has advantages and disadvantages. And Java has just historically always tried to "hide complexity", which in practice however paradoxically multiplies complexity _if youre not already used to the pattern used_.
Yeah, I remember the first time I encountered a spring project (well before boot was out) and just about lost my shit with how much magic was happening.
It is productive once you know a whole lot about it though, and I already had to make that investment so might as well reap the rewards.
Instead, official documentation seems comfortable with recommending a third party package: https://docs.python.org/3/library/urllib.request.html#module...
>The Requests package is recommended for a higher-level HTTP client interface.
Which was fine when requests were the de-facto-standard only player in town, but at some point modern problems (async, http2) required modern solutions (httpx) and thus ecosystem fragmentation began.
The h11, h2, httpcore stack is probably the closest thing to what the Python stdlib should look like to end the fragmentation but it would be a huge undertaking for the core devs.
More importantly, it would be massively breaking to remove the existing functionality (and everyone would ignore a deprecation), and confusing not to (much like it was when 2.x had both "urllib" and "urllib2").
It'd be nice to have something high level in the standard library based on urllib primitives. Offering competition to those, not so much.
The stdlib may not be the best, but the fact all HTTP libs that matter are compatible with net/http is great for DX and the ecosystem at large.
It is not built for convenience. It has no methods for simply posting JSON, or marshaling a JSON response from a body automatically, no "fluent" interface, no automatic method for dealing with querystring parameters in a URL, no direct integration with any particular authentication/authorization scheme (other than Basic Authentication, which is part of the protocol). It only accepts streams for request bodys and only yields streams for response bodies, and while this is absolutely correct for a low-level library and any "request" library that mandates strings with no ability to stream in either direction is objectively wrong, it is a rather nice feature to have available when you know the request or response is going to be small. And so on and so on.
There's a lot of libraries you can grab that will fix this, if you care, everything from clones of the request library, to libraries designed explicitly to handle scraping cases, and so on. And that is in some sense also exactly why the net/http client is designed the way it is. It's designed to be in the standard library, where it can be indefinitely supported because it just reflects the protocol as directly as possible, and whatever whims of fate or fashion roll through the developer community as to the best way to make web requests may be now or in the future, those things can build on the solid foundation of net/http's Request and Response values.
Python is in fact a pretty good demonstration of the risks of trying to go too "high level" in such a client in the standard library.
Yes, and it's in the standard library (System namespace). Being Microsoft they've if anything over-featured it.
And while this article [1] says "It's been around for a while", it was only added in .NET Framework 4.5, which shows it took a while for the API to stabilise. There were other ways to make web requests before that of course, and also part of the standard library, and it's never been "difficult" to do so, but there is a history prior to HttpClient of changing ways to do requests.
For modern dotnet however it's all pretty much a solved problem, and there's only ever been HttpClient and a fairly consistent story of how to use it.
[1] https://learn.microsoft.com/en-us/dotnet/core/extensions/htt...
is 14 years not a while?
.NET’s WebRequest was available in .NET Framework 1.1 in 2003 [1].
But since then, Microsoft noticed the issues with WebRequest and came up with HttpClient in 2012. It has some issues and footguns, like those related to HttpClient lifetime, but it’s a solid library. On the other hand, the requests library for Python started in 2011 [2], but the stdlib library hasn’t seen many improvements.
[0] https://github.com/python/cpython/blob/6d7e47b8ea1b8cf82927d...
[1] https://learn.microsoft.com/en-us/dotnet/api/system.net.webr...
[2] https://github.com/psf/requests/blob/main/HISTORY.md#001-201...
it's called the STD lib for a reason...
const response = await new Promise( (resolve, reject) => {
const req = https.request(url, {
}, res => {
let body = "";
res.on("data", data => {
body += data;
});
res.on('end', () => {
resolve(body);
});
});
req.end();
});> Undici is an HTTP client library that powers the fetch API in Node.js. It was written from scratch and does not rely on the built-in HTTP client in Node.js. It includes a number of features that make it a good choice for high-performance applications.
I've often ended up reimplementing what I need because the API from the famous libraries aren't efficient. In general I'd love to send a million of requests all in the same packet and get the replies. No need to wait for the first reply to send the 2nd request and so on. They can all be on the same TCP packet but I have never met a library that lets me do that.
So for example while http3 should be more efficient and faster, since no library I've tried let me do this, I ended up using HTTP1.1 as usual and being faster as a result.
Python makes everything so easy.
I realized this the other day, and dub it Bram's Law -- Bram
Bram's Law
The easier a piece of software is to write, the worse it's implemented in practice. Why? Easy software projects can be done by almost any random person, so they are. It's possible to try to nudge your way into being the standard for an easy thing based on technical merit, but that's rather like trying to become a hollywood star based on talent and hard work. You're much better off trading it all in for a good dose of luck.
This is why HTTP is a mess while transaction engines are rock solid. Almost any programmer can do a mediocre but workable job of extending HTTP, (and boy, have they,) but most people can't write a transaction engine which even functions. The result is that very few transaction engines are written, almost all of them by very good programmers, and the few which aren't up to par tend to be really bad and hardly get used. HTTP, on the other hand, has all kinds of random people hacking on it, as a result of which Python has a 'fully http 1.1 compliant http library which raises assertion failures during normal operation.
Remember this next time you're cursing some ubiquitous but awful third party library and thinking of writing a replacement. With enough coal, even a large diamond is unlikely to be the first thing picked up. Save your efforts for more difficult problems where you can make a difference. The simple problems will continue to be dealt with incompetently. It sucks, but we'll waste a lot less time if we learn to accept this fact.
I've noticed that many languages struggle with HTTP in the standard library, even if the rest of the stdlib is great. I think it's just difficult to strike the right balance between "easy to use" and "covers every use case", with most erring (justifiably) toward the latter.
I still hear people complain about how such and such removal between "minor versions" of Python 3 (you really should be thinking of them as major versions nowadays — "Python 3 is the brand", the saying goes now), where they were warned like two years in advance about individual functions, supposedly caused a huge problem for them. It's hard for me to reconcile with the rhetoric I've heard in internal discussions; they're so worried in general about possible theoretical compatibility breaks that it seems impossible to change anything.
The notable exception is Go, which has a fantastic one. But Go is pretty notable for having an incredible standard library in general.
1. Whether or not it was "made for" webdev, people do use Rust for that.
2. Plenty of people write networked Rust that interacts with HTTP. That code requires an HTTP stack, even if it isn't web development.
3. Like all of the other examples, Rust does have an excellent third-party HTTP stack (reqwest and its underpinnings). So it's not like Rust fails to do HTTP.
(As an outsider I had the impression that Go's net/http was good, but a lot of people in this thread are complaining about it as well. So it may be 0-4 instead of 1-3).
Node.js got its production version in 2023.
Rust doesn't include an HTTP client at all.
Even for stdlib that have a client, virtually none support HTTP/3, which is used for 30% of web traffic. [1]
--
HTTP (particularly 2+) is a complex protocol, with no single correct answers for high-level and low-level needs.
Always remember that open-source is an author’s gift to the world, and the author doesn’t owe anything to anyone. Thus, if you need a feature that for whatever reason can’t or won’t go upstream, forking is just about the only viable option. Fingers crossed!
Put your side project on your personal homepage and walk away - fine.
Make it central infrastructure - respond to participants or extend or cede maintainership.
FOSS means the right to use and fork. That's all it means. That's all it ever meant. Any social expectations beyond that live entirely in your imagination.
There is simply no responsibility an OSS maintainer has. They can choose to be responsible, but no one can force them. Eventually OSS licensing is THE solution at heart to solve this problem. Maintainers go rogue? Fork and move on. But surprise, who is going to fork AND maintain? Filling in all the demands from the community, for potentially no benefit?
No one can force him to take the responsibility, just like no one can force anyone else to.
This doesn't come over night and this is a spectrum and a choice. From purely personal side project over exotic Debian package to friggin httpx with 15k Github stars and 100 million downloads a week the 46th most downloaded PyPI package!
If this shall work reasonably in any way, hou have to step up. Take money (as they do, https://github.com/sponsors/encode), search fellow maintainers or cede involvement - even if only temporarily.
An example of a recent, successful transition is UniGetUI https://github.com/Devolutions/UniGetUI/discussions/4444
I feel there should be support from the ecosystem to help with that. OpenJS Foundation seems doing great: https://openjsf.org/projects. The Python Software Foundation could not only host PyPI but offer assistance for the most important packages.
This is an oxymoron. Either you have obligations, or you don't. There's no such thing as having "no obligations" but also "growing responsibility".
I don't understand how you can possibly conclude that just because you've chosen to become dependent on some FOSS library, they owe you anything. You don't get to somehow impose obligations on other people by your choices. They get none of your profits, but they're somehow responsible to you for your business risks? Nonsense.
It is a condition of your use of the code that you've accepted its license, and FOSS licenses are CRYSTAL CLEAR (ALL CAPS) on what obligations or responsibilities the authors have towards you - none whatsoever. Your use of the software is contingent on your acceptance of that license.
If that lack of warranty poses an unacceptable business risk to you, go buy support. Pay a dev to fix the issues you're having, rather than inventing some fictitious responsibility they have to you to do it for free.
There is also the possibility that a maintainer simply doesn't care about what the community wants, it's his baby and he can do what he wants.
Forking a project is built-in by licensing. A lot of complaints, but those complainers don't fork. Why is that? Yeah right.
Side Note: Transferring projects to foundations etc with funding may be a solution for projects that are highly depended on and require active, reliable maintenance. They wont work well for innovation or experimentation. Just saying they are just a part of the equation and not the sole solution.
If you need stronger guarantees, pay someone to deliver them.
One of its intended use cases is bridging contribution gaps: while contributing upstream is ideal, maintainers may be slow to merge contributions for various reasons. Forking in response creates a permanent schism and a significant maintenance burden for what might be a small change. Modshim would allow you to create a new Python package containing only the fixes for your bugbears, while automatically inheriting the rest from upstream httpx.
There are a couple of example of this readme: (1) modifing the TextWrapper object but then use it through the textwrap library's wrap() function, and (2) modifing the requests Session object, but then just using the standard requests.get(). Without modshim (using standard monkey-patching) you would have to re-implement the wrap and get methods in order to bind the new TextWrapper / Session classes.
- what's the performance like for big packages (say, pytorch)? Have you done some benchmarking?
- is typing kept for the shims? My immediate guess, again, without even trying it out, is not. If yes, how?
edit: formatting
"So what is the plan now?" - "Move a little faster and not break things"
Loved that little detail, reminds me of the old interwebs :)
- httpx
- curl cffi
- httpmorph
- httpcloak
- stealth crawler
I wrote a framework, link below, which uses them all. You can compare each to verify crawling speed. Some sites can be cleanly crawled with a one particular framework.
Having read the article I am in a pain. I do break things while development. I rewrite stuff. Maybe some day I will find a way to develop things "stable". One thing I try to keep in good shape is 'docker' image. I update it once everything seems to be quite stable.
What uv does is parallelize the final download of packages after resolution, and batch pre-fetch metadata during resolution. I don't think these benefit from async, due to their batch nature classic multi-threaded download pools are probably the better solution, but I could be wrong!
Experiments have been done on the former in pip and didn't find much/any improvement in CPython, this may change in free threaded CPython. For the latter we currently don't have the information from the resolver to extract a range of possible metadata versions we could pre-range, I am working on this but it requires new APIs in packaging (the Python library) and changes to the resolver, and again we will need to benchmark to see if adding pre-fetching actually improves things.
This certainly wouldn't be the first time an author of a popular library got a little too distracted on the sequel to their library that the current users are left to languish a bit.
Just a small headsup: clicking on the Leiden Python link in your About Me page give not the expected results.
And a small nitpick: it's "Michiel's" in English (where it's "Michiels" in Dutch).
Thanks for devoting time to opensource... <3
Or maybe it is that your brain is cooked already, or is on the brink, and your condition attracts you to HTTP and Python, after which it basically has you.
The only way to not go bonkers is to design a library by commitee, so that the disease spreads evenly and doesn't hit any one individual with full force. The result will be ugly, but hopefully free of drama.
Whether actively defending your trademark is actually required is a bit of a nuanced topic. Generally, trademarks can be lost through genericide (the mark becomes a generic term for the type of product) or abandonment. Abandonment happens when either the mark owner stops using the mark itself, or takes an action that weakens the mark. The question, then, is whether failing to defend infringing use constitutes a weakening action. Courts differ on this, and there is a large gray area between "we didn't immediately sue a local mom-and-pop shop" and "we allowed a rival company to use the mark erroneously across several states for years without taking action."
I think if had named it HTTPX2 or HTTPY, that would be much worse because it asserts superiority without earning it. But he didn't.