[RonnyPfannschmidt]

Requested by Ronny · project thread

🤖 Written by Claude Opus 5.5 via Claude Code for the pytest maintainers; I prompted it, it did the work, I read it.

Closes #15254.

TempPathFactory.mktemp() and tmp_path used make_numbered_dir(), which lists the whole basetemp on every call to find the next free number. basetemp gains entries with every test (two directories and a *current symlink per test in the issue's reproducer), so a session does O(N²) work. At N=3000, profiling put 41 of 54 s in find_prefixed.

This PR is a simpler alternative to the stat-validated listing cache proposed in the issue. The factory keeps a per-prefix counter (_next_number, keyed by the lower-cased prefix):

• mktemp(numbered=True), and through it tmp_path and pytester, counts from 0 without listing basetemp. The factory always creates basetemp itself (a fresh pytest-N, or a --basetemp it has just wiped).
• mkdir decides uniqueness. A name it fails on (a directory, file or dangling symlink already there, or a PermissionError on Windows for a name still being deleted) is skipped. After 10 failures in a row, basetemp is listed once to jump past the run. After 20, an OSError chained to the last error is raised, so a filesystem that reports every name as taken cannot make it loop forever.
• make_numbered_dir() and its other callers (the pytest-of-<user> session root) are unchanged.

Behaviour differences

• mktemp numbers are no longer "greater than any existing one". Existing names are skipped instead, and the docstring now says so. With a basetemp the factory owns, that only matters for directories something else creates there.
• Numbers never go down. With main, a number freed by the tmp_path retention removal could be reused.
• On a case-sensitive filesystem, a foreign X2 no longer pushes x past 2, because mkdir decides the collision. Main matched the listing case-insensitively.

Covered by tests: no listing in mktemp, unexpected entries (directory, file, dangling symlink), a run of 50 taken names (one listing), overlapping prefixes (test_a and test_a1 both produce test_a10), the probing limit, and retries on other OSErrors.

Numbers

Reproducer from the issue (pytest -q -p no:cacheprovider, noisy cloud box, wall-clock):

tests	main	this PR
1000	3.9 s	~3–4 s
2000	10.9 s	~3.5 s
4000	37.3 s	~7.5 s

Run time now grows linearly. What remains is mkdir/symlink syscalls.

Checklist

• Include new tests or update existing tests when applicable.
• Allow maintainers to push and squash when merging my commits.
• Add text like closes #XYZW to the PR description and/or commits.
• If AI agents were used, they are credited in Co-authored-by commit trailers.
• Create a new changelog file in the changelog directory.

Full test suite passed locally (4662 passed). pre-commit is clean.

[nicoddemus]

Looks good, it seems this fixes the original issue, but left a comment for your consideration to further improve things.