Take pytest 9.1.2 when it ships: #73's exclusion of 9.1.0/9.1.1 #83

Activity

1. gregoryfoster commented on Oct 28, 2026

   

   gregoryfoster

   on Oct 28, 2026

   MemberAuthor

   More actions

   Hand-off brief: #83 (coordinator, #80, 2026-10-28)

   Unblocked by pytest 9.2.0, not 9.1.2. 9.2.0 went up on PyPI 2026-10-27T21:30Z. 9.1.2 never shipped: PyPI goes from 9.1.1 (2026-06-19) straight to 9.2.0. The 9.2.0 release notes list pytest-dev/pytest#14635: "Fixed fixture resolution failures when specifying multiple test paths on the command line that share common parent directories". That's #73's cause. pyproject.toml's spec pytest>=9.0.3,!=9.1.0,!=9.1.1,<10 already admits it, so the issue's recipe works as written, with 9.2.0 in place of 9.1.2.

   9.2.0 is a minor release: check its breaking changes against this repo

   From the 9.2.0 notes, Removals and backward incompatible breaking changes:

   • -c with a missing, directory or unloadable path is now a usage error (-c/--config-file has two failure modes for invalid paths: silent no-op for unrecognized extensions, and raw FileNotFoundError for recognized ones pytest-dev/pytest#14716). Grep scripts/, .githooks/, CI and the tests for a pytest -c.
   • parametrize's indirect, ids and scope are keyword-only (Improve error output when specifying multiple argument strings with parametrize pytest-dev/pytest#8593). Grep for positional uses.
   • Plugin import failures exit 4 or 5, not 1 (ImportError raised by plugins make pytest return EXIT_TESTSFAILED pytest-dev/pytest#993). addopts loads -p pytester. Check that nothing reads pytest's exit code as 1 for a plugin error, such as tests/deploy/ tests that spawn pytest through pytester, or test_pytest_collection.py's orders.
   • Read the rest of the notes, including the deprecations, against filterwarnings in pyproject.toml: a new deprecation warning there becomes a failure.

   Recipe

   uv lock --upgrade-package pytest, then uv sync --group dev, then uv run pytest tests/deploy/test_pytest_collection.py (all six orders), then the full suite, then ruff. A red collection test means 9.2.0 lacks the fix after all: say so and stop.

   Decide

   • Keep or drop the !=9.1.0,!=9.1.1 exclusions and their comment. The issue allows either. With 9.1.2 never shipping, a floor of >=9.2.0 would say the same more plainly, and its comment would point at Error collecting tests on 9.1.1 - fixtures not found when a parent directory appears multiple times in a argument list  pytest-dev/pytest#14635 and Targeted pytest runs that interleave tests/ and tests/deploy/ modules lose tests/deploy/conftest.py's fixtures #73.
   • The comment at pyproject.toml lines 40 to 43 names 9.1.2. Correct it whichever way you go.

   Shipping

   pyproject.toml and uv.lock are runtime paths (drift.py's RUNTIME_FILES), even though pytest is dev-only. Ship with scripts/deploy.sh on origin/main's tip, detached, with the operator's approval, the same as #73 and #124. Before it, read readlink /srv/broker/live and journalctl -t broker-deploy -n 5.

   Verify

   • CI is green by job, audit included, on the final SHA.
   • uv.lock holds pytest 9.2.0.
   • The node's tests/deploy and full-suite counts match the baseline, with the same 4 skips.
   • journalctl -t broker-deploy ends with live is on <sha12>, and drift reads ok.

   Baseline [measured 2026-10-28T05:4xZ]

   • live -> releases/714db2740140; drift "live 714db27 is main".
   • ACL LOG: 5.
   • CI and audit green by job on 714db27.
   • Node tests/deploy 801 passed, 4 skipped; full suite 1348 passed, 4 skipped (both with /etc/broker/.env sourced; without it, tests/deploy skips 21). Adopt the cohort disk GC on co-broker (gregoryfoster/skills maintaining-host-disk); retire prune_editor_builds.sh #127's retired test file took 6 off each.

   Out of scope

   Other dependency bumps, Dependabot's config, and pytest-asyncio unless 9.2.0 forces it (say so if it does).

   Ship

   Close #83 with the lock change, the six orders' result, what each breaking change turned out to touch, the exclusions decision, the commit range, and the deploy evidence. Don't tick #80. Then return to #80's coordinator.
2. gregoryfoster commented on Oct 28, 2026

   

   gregoryfoster

   on Oct 28, 2026

   MemberAuthor

   More actions

   Done: lock on pytest 9.2.0 (88976b8)

   9.1.2 never shipped; 9.2.0 (PyPI 2026-10-27T21:30Z) is the first release with pytest-dev/pytest#14635's fix, per its notes.

   Lock. uv.lock: pytest 9.0.3 -> 9.2.0, nothing else moved (pluggy already 1.6.0, which 9.2.0 now requires). pytest-asyncio untouched (1.4.0); 9.2.0 forced nothing there.

   Six orders. tests/deploy/test_pytest_collection.py: 6 passed on 9.2.0, including both interleaving orders (test_a-test_top-test_b, test_b-test_top-test_a).

   Breaking changes vs this repo (9.2.0 Removals and backward incompatible breaking changes):

   • -c usage error (-c/--config-file has two failure modes for invalid paths: silent no-op for unrecognized extensions, and raw FileNotFoundError for recognized ones pytest-dev/pytest#14716): touches nothing. No pytest -c in scripts/, .githooks/, .github/, tests/, src/, deploy/.
   • Keyword-only indirect/ids/scope (Improve error output when specifying multiple argument strings with parametrize pytest-dev/pytest#8593): touches nothing. An AST scan of 110 parametrize calls under tests/ found none with more than two positional args.
   • Plugin import failures exit 3/4, not 1 (ImportError raised by plugins make pytest return EXIT_TESTSFAILED pytest-dev/pytest#993): touches nothing. -p pytester loads. Nothing reads pytest's exit code for a plugin error: test_pytest_collection.py uses assert_outcomes, and test_pytest_tmp_retention.py's rc is the inner session's test-failure exit. The returncode == 1 asserts in test_deploy.py are deploy.sh's.
   • Deprecations vs error::DeprecationWarning: none fire. No yield_fixture, no CallSpec2, no write_line. The one fixture defined as a method (test_shellcheck_gate.py's TestTheHook.clone) is function-scoped, outside Extend instance method fixture deprecation to all higher scopes pytest-dev/pytest#14774. Under _getini_ini: "string" type returns list silently instead of raising TypeError pytest-dev/pytest#14808, tmp_path_retention_count = 1 is an int, which Native TOML numeric faulthandler_timeout is rejected pytest-dev/pytest#15132 now accepts natively. No warning, and the suite is green under the filter.

   Exclusions: dropped for a floor. pytest>=9.0.3,!=9.1.0,!=9.1.1,<10 -> pytest>=9.2.0,<10. With 9.1.2 never shipping, a floor says it more plainly. The pyproject.toml comment now names 9.2.0 and pytest-dev/pytest#14635 / #73 instead of 9.1.2. The collection test's docstring dropped "unreleased" and now points at the floor.

   Commits. 6a4cc80..88976b8 (one commit, 88976b8).

   CI on 88976b8, green by job: CI lint + test (40719971136), uv audit (report-only) (40719971454).

   Node counts (~/broker on 88976b8, /etc/broker/.env sourced): tests/deploy 801 passed, 4 skipped; full suite 1348 passed, 4 skipped. Both match the 714db27 baseline, with the same 4 skips.

   Deploy (scripts/deploy.sh 88976b8, detached, operator-approved). Before: live -> releases/714db2740140, journal ended live is on 714db2740140 (2026-10-27T23:05:23Z). After, from journalctl -t broker-deploy:

   2026-10-28T16:15:54Z CI passed for 88976b8a21f5
   2026-10-28T16:15:58Z live -> 88976b8a21f5 (was releases/714db2740140)
   2026-10-28T16:16:00Z the backup run on 88976b8a21f5: unchanged
   2026-10-28T16:20:40Z the probe's tick 43aec5ecbd6d4ba38e12d67e828e09b9 on 88976b8a21f5: exited 0
   2026-10-28T16:20:40Z live is on 88976b8a21f5
   

   Drift at 2026-10-28T16:45:00Z: ok, "live 88976b8 is main" (checkin 202).

   ACL LOG: 5 entries, unchanged. No denial was caused.
3. gregoryfoster

   closed this as completedon Oct 28, 2026