Make async dependencies optional. #2858
Conversation
|
Tests currently blocked on |
Resolved |
| http2 = [ | ||
| "h2>=3,<5", | ||
| "httpcore[http2]", | ||
| ] | ||
| socks = [ | ||
| "socksio==1.*", | ||
| "httpcore[socks]", | ||
| ] | ||
| asyncio = [ | ||
| "httpcore[asyncio]" | ||
| ] | ||
| trio = [ | ||
| "httpcore[trio]" | ||
| ] |
There was a problem hiding this comment.
This is interesting... currently we're including the httpcore dependencies explicitly.
Eg. pip install httpx[socks] uses the dependency socksio==1.* instead of the dependency httpcore[socks].
The second style makes more sense, because that's what we're actually trying to effect... but I'm a bit hesitant that some Python install tooling might incorrectly fail to install httpcore[socks] as an optional dependency when the httpcore dependancy is already included. Is my hesitancy well founded here, or are we okay to use this new style throughout?
There was a problem hiding this comment.
I think this should always work but you're right that it's tool dependent and the Python specification around this does not provide guarantees.
Co-authored-by: T-256 <132141463+T-256@users.noreply.github.com>
|
remove anyio. |
|
If async is optional, should we improve the unit test so that it can skip some tests when certain optional dependencies are not installed? And can we consider conducting separate tests in CI for both scenarios: when optional dependencies are not installed and when they are already installed? |
|
@abersheeran I don't know. We don't have any need for that in CI, but perhaps some folks would appreciate fast runs in local dev. Perhaps spin it out into a new discussion? |
It's probably worthwhile to make sure the project works as intended in some basic form during CI without the optional dependencies. |
|
I'm not sure it's fair to say we should have a separate discussion when we need something like it for reasonable coverage of the changes introduced in this pull request. I think "skip tests when the dependencies are not available for better developer experience" makes sense as a new discussion, but "ensure the library works without the async dependencies installed" is critical here. Otherwise, we're just relying on manual testing to ensure the default installation of the library isn't broken. If you have a strong preference for a follow-up pull request, I can get behind that, but I think we need automated testing before release. |
|
@zanieb Ah interesting. I've perhaps been missing your intent here. I had "should we ensure the tests can run without all dependencies installed?". |
|
👍
Yeah I think this is the idea. We should run some subset of the tests (i.e. the synchronous API) in CI with the default install dependencies to ensure things are working as intended now (and in the future). |
|
@zanieb Do we want to block this on "run the tests without all dependencies installed" or not? There's a chunk of work there that I'd suggest would need to be taken on incrementally if we want to deal with it. |
I'm okay with merging this and addressing as a follow-up! |
Co-authored-by: Zanie Blue <contact@zanie.dev>
Co-authored-by: T-256 <132141463+T-256@users.noreply.github.com>
See the "Version 1.0, Minimising project dependancies" discussion for background, as well as the proposed PR for
httpcoreversion 1.0.This pull request would changes our default install footprint from...
httpcorecertifih11sniffioanyioidnaTo this...
httpcorecertifih11idnaWe do also have scope to make
idnasupport optional, but this pull request doesn't address that.