Load each config file and probe each candidate only once in UpwardSearch

[alexander-beedie]

Summary

Overview

UpwardSearch was doing unnecessary work; in most packages it wasn't a problem, but in some it could trigger a lot of redundant filesystem work by repeating a lot of load/stat calls on the same file(s).

Solution

• Optimise-out redundant IO/syscalls from UpwardSearch (config load/probes) by using simple SmallMap caching, keyed by path.

• Before this PR, one config search would load the same file and stat the same candidate paths multiple times; now it only does each once.

Benchmarks¹

There's a nice positive impact (on a limited number of packages), but it can be even more significant in corporate environments running endpoint security software.

Note: I was finally able to deterministically reproduce the increased cost of syscalls/IO caused by corporate endpoint security software introspection, so I've included it in the timings below (refs: #3993, #4017 (comment)).

• The "taxed" columns show the additional overhead from endpoint security software. The real standout is pywin32, with a ~5-6x overhead: ~0.74s normally, ~4.29 secs with the security software tax (reduced to ~2.83 secs with this PR).

  Package	Speedup (s)	Δ	Taxed speedup (s)	Taxed Δ	Tax
  vision	0.54 → 0.36	-32.50%	0.54 → 0.37	-31.70%	1.0×
  materialize	1.95 → 1.48	-23.30%	1.97 → 1.48	-24.80%	1.0×
  kornia	0.55 → 0.45	-19.10%	0.58 → 0.46	-20.10%	1.0×
  pycryptodome	0.28 → 0.24	-15.70%	0.29 → 0.24	-15.90%	1.0×
  beartype	0.45 → 0.38	-15.60%	0.47 → 0.39	-14.70%	1.0×
  mitmproxy	0.57 → 0.48	-14.90%	0.60 → 0.52	-13.60%	1.1×
  graphql-core	0.38 → 0.33	-13.80%	0.38 → 0.33	-12.20%	1.0×
  poetry	0.91 → 0.81	-10.90%	1.34 → 1.06	-21.10%	1.5×
  spack	0.55 → 0.50	-9.30%	0.58 → 0.53	-8.60%	1.1×
  pip	0.66 → 0.60	-8.90%	0.63 → 0.59	-5.60%	1.0×
  core	4.14 → 4.02	-3.40%	5.26 → 4.86	-7.30%	1.3×
  psycopg	0.65 → 0.63	-2.80%	1.88 → 1.36	-27.60%	2.9×
  pylox	0.56 → 0.55	-2.50%	1.77 → 1.27	-28.10%	3.2×
  archinstall	0.76 → 0.75	-2.40%	1.46 → 1.43	-3.40%	1.9×
  manticore	0.67 → 0.66	-2.20%	1.73 → 1.21	-27.50%	2.6×
  pywin32	0.74 → 0.72	-2.00%	4.29 → 2.83	-33.30%	5.8×
  dd-trace-py	1.63 → 1.57	-2.00%	2.25 → 1.99	-4.80%	1.4×
  narwhals	1.05 → 1.04	-2.00%	2.67 → 2.57	-4.30%	2.5×
  scipy	1.08 → 1.04	-2.00%	1.95 → 1.81	-3.40%	1.8×

• Speedups are almost entirely wall-clock, not CPU(load/stat costs are IO, not compute).

• Projects with no .venv benefit most, as they have to discover their own interpreter, triggering redundant config loads. The "taxed" winners are projects with a large site-packages tree which can trigger more syscalls.

Test Plan

• All existing unit tests pass, one new test added.

• Was able to trigger the endpoint security cost via symlink indirection to the test corpus (eg: on macOS /tmp is a symlink to /private/tmp, so /tmp/primer_repos/… crosses one symlink and /private/tmp/primer_repos/… crosses none; identical files/inodes/bytes, but it's the difference between a syscall costing ~0.7 µs or ~1.28 ms (~1000x more expensive, ouch 🚫)

Footnotes

1. Test machine: Apple Silicon M3 Max (16 cores). ↩

[meta-codesync]

This pull request has been imported. If you are a Meta employee, you can view this in D116264606. (Because this pull request was imported automatically, there will not be any future comments.)