Do CLI import in __main__, not __init__ #3547
Conversation
88baca0 to
cccc8cb
Compare
| @@ -59,7 +44,6 @@ def main() -> None: # type: ignore | |||
| "InvalidURL", | |||
| "Limits", | |||
| "LocalProtocolError", | |||
| "main", | |||
There was a problem hiding this comment.
If assume this causes breaking change (deletion from public api) we may need to keep it as pointer to __main__.main which also lazy import it:
def main():
from .__main__ import main as orignial_main
original_main()by this way, there is no need to update test.
There was a problem hiding this comment.
good point, will make this change
There was a problem hiding this comment.
Turns out this is a little more complicated, unfortunately. The main function in httpx/_main.py is a click.BaseCommand, not just a typing.Callable. Wrapping it in a new def main(...) still changes the public API, since httpx.main is no longer a click.BaseCommand. This gets exposed by the tests, which attempt to access attributes of a click.BaseCommand object that aren't there for a simple function. To address, I believe we'd have to have httpx.__init__.py import click, which leads to the issue we're trying to get around in the first place, or we'd have to do some funky mocking magic to make main look more like a click.BaseCommand. I'm happy to explore that if that's the direction you'd like this to head, but doesn't feel quite right to me at this point.
In summary, I unfortunately can't think of a good way right now to maintain the current public API and achieve the goal of this PR.
There was a problem hiding this comment.
Wrapping it in a new
def main(...)still changes the public API
Also, it could use PEP562 (module level __getattr__) instead of wrapping:
def __getattr__(name):
if name == "main":
from .__main__ import main
return main
raise AttributeError(f"module {__name__!r} has no attribute {name!r}")There was a problem hiding this comment.
Thanks, updated. Removed the __main__.py module to keep the diff as small as possible. Can add back, or it can go in a separate PR since it's a separate feature.
28a382b to
fab0285
Compare
Importing the CLI code doubles the import time for httpx altogether. By importing lazily at runtime, the additional time only happens when running the httpx CLI, and not on every "import httpx".
Summary
From #3523
Importing the
_main.pymodule inhttpx/__init__.pyadds 0.087s worth of import time on my machine (MacBook M1 Pro, 32 GB RAM, macOS 15.3.1) running Python 3.13.2.It's certainly not a huge amount of time in the absolute, but it roughly doubles the amount of time it takes to
import httpx, even when you're not intending to run the CLI. Here's an import time graph generated using tuna on the output ofpython -X importtime -c "import httpx" > import.logThis PR adds
httpx/__main__.pythat does the import instead. This also allowspython -m httpxas in #3216.It does change the public API:
httpx.mainbecomes the (more ungainly)httpx.__main__.main.Checklist