Speed up import time by deferring inspect

[danielhollas]

Deferring import of inspect cuts the import time by ~10% (4ms on my machine).

CPython issue: python/cpython#118761

I haven't been able to run the tests locally, seeing these errors:

RROR tests/test_api.py - AttributeError: module 'tests.fixtures' has no attribute 'EggInfoPkg'
ERROR tests/test_integration.py - AttributeError: module 'tests.fixtures' has no attribute 'DistInfoPkg'
ERROR tests/test_main.py
ERROR tests/test_zip.py - AttributeError: module 'tests.fixtures' has no attribute 'ZipFixtures'

Benchmarks

These have been run with latest CPython main branch (as of Aug 6th 2024), these gains are likely representative for Python 3.13, but not 3.12.

this PR

❯ hyperfine -w 5 'python -c "import importlib_metadata"'
Benchmark 1: python -c "import importlib_metadata"
  Time (mean ± σ):      39.6 ms ±   4.6 ms    [User: 29.8 ms, System: 9.4 ms]
  Range (min … max):    29.7 ms …  49.8 ms    68 runs

main

❯ hyperfine -w 5 'python -c "import importlib_metadata"'
Benchmark 1: python -c "import importlib_metadata"
  Time (mean ± σ):      43.4 ms ±   3.6 ms    [User: 33.1 ms, System: 9.8 ms]
  Range (min … max):    34.9 ms …  50.1 ms    68 runs

[danielhollas]

This is a trick that is used in stdlib. I am not sure it is needed here, but since get_toplevel_name is called in a loop from _top_level_inferred perhaps it is warranted to avoid the overhead of calling import inspect repeatedly.

[jaraco]

I'm not a fan of global variables or the additional complexity. This change introduces enough disruption to the essential logical flow that I'm -1. Do we know how much overhead there is in repeated import inspect? My understanding (which may be incorrect) is that import inspect is essentially a dict lookup if it's already been imported. I'm guessing the overhead is acceptable. Can we try a simple deferral for now?

[jaraco]

Thanks for the contrib!

[jaraco]

I'm not a fan of global variables or the additional complexity. This change introduces enough disruption to the essential logical flow that I'm -1. Do we know how much overhead there is in repeated import inspect? My understanding (which may be incorrect) is that import inspect is essentially a dict lookup if it's already been imported. I'm guessing the overhead is acceptable. Can we try a simple deferral for now?

[jaraco]

Since we don't have a test protecting this behavior, I'd like to see a comment pointing to the issue, so a future someone doesn't refactor this optimization away.

[jaraco]

I've minimized the change to address the specific issue. I'll deal with the typeshed ignore workaround separately.

[jaraco]

I should have added a news fragment before merging. I added it later in 71b4678.

[jaraco]

This change is released in v8.4.0.