Refs #34118 -- Avoided repeat coroutine checks in MiddlewareMixin.

[adamchainz]

32d70b2 changed MiddlewareMixin to call iscoroutinefunction on every request. Whilst fast, it’s not free, with many function calls and unwrapping in its tree: https://github.com/python/cpython/blob/88bac5d5044e577825db1f9367af908dc9a3ad82/Lib/inspect.py#L420 . Multiplying this by many middlewares it will add up.

This PR changes to cache the result in an instance variable, as it did before.

Quick benchmark shows iscoroutinefunction can take ~0.5us on a function with one decorator:

Python 3.12.0 (main, Oct  5 2023, 12:02:11) [Clang 15.0.0 (clang-1500.0.40.1)]
Type 'copyright', 'credits' or 'license' for more information
IPython 8.16.1 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import inspect

In [2]: def f(): pass

In [3]: %timeit inspect.iscoroutinefunction(f)
358 ns ± 1.62 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

In [4]: from functools import cache

In [5]: @cache
   ...: def g(): pass

In [6]: %timeit inspect.iscoroutinefunction(g)
544 ns ± 1.22 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

[carltongibson]

Yes, this seems totally reasonable to me.

(The asgiref.sync utilities include a check that they're correctly passed a coroutine marked class, so an error here would show in the test output, I trust.)

[felixxm]

@adamchainz Thanks 👍

[adamchainz]

Thanks! Look after the microseconds and the milliseconds will look after themselves 😉

[adamchainz]

@felixxm backport so 5.0 doesn't have the performance regression?

[felixxm]

@felixxm backport so 5.0 doesn't have the performance regression?

I categorized this as a cleanup. If we want to treat this as a bug it should be a release blocker for Django 4.2, so a separate ticket and a release note in 4.2.7.txt would be needed.

[adamchainz]

Oh sorry, I didn't realize that commit was in 4.2, I thought it was only on the 5.0 branch.

Eh, I can't be bothered filing a ticket if no one else has spotted this.

[carltongibson]

(I don't think it's release blocker status...)