[OffgridwithJD]

Part 1 of #2038.

What changes

The nightly suites job runs PG 18 on aarch64 as well as x86_64. Its own comment gives the reason: a missing barrier in concurrent code is invisible under x86_64's stronger store ordering, and a defect on aarch64. Nightly pytest-cluster ran on x86_64 only, so on aarch64 only the shell suites executed. That is one of the three things keeping the shell suites from being retired.

• nightly.yml: pytest-cluster takes the suites job's legs as an include: matrix, adds { pg: '18', runner: ubuntu-24.04-arm }, and puts the architecture in the job name, as suites does. The apt cache keys already carry runner.arch. Every pinned wheel in requirements-test.txt (pytest, xdist, psycopg[binary], pyarrow, duckdb) is published for manylinux aarch64.
• test_pytest_runs_where_the_suites_run.py (new guard) derives the property: in ci.yml and nightly.yml, every (major, architecture) the suites job runs, pytest-cluster runs too. It reads include: entries, pg:/runner: lists, and a literal runs-on:. A workflow it cannot compare is named (unreadable, no suites legs, or no pytest-cluster job), so an empty answer means every leg is covered.

Evidence

• Red, before the workflow change: pytest-cluster runs every (major, architecture) the suites job runs: got 'nightly.yml:18/aarch64' want 'none'.
• Green after it: 2 passed. The reader arm also proves each matrix form and each "cannot compare" case on fixtures.
• Push gate: guards 865 passed; collected 5347, matching cluster_tests; ledger census 4894 agrees; gate rc 0; harness_selftest passed.

Not proven here

The aarch64 leg itself runs only in the nightly, so the first nightly after merge is its first run. The x86_64 legs are unchanged. If the arm run needs more than the 40-minute limit, I'll follow up with a per-leg timeout.

Registries

• NO_CLUSTER gains the new file.
• guard_tests 866 → 868, with a comment line.
• TESTS.md §357, restated onto main after test(pytest): port extension_upgrade.sh, and make the parity tooling see standalone gates (#2032) #2036. The sanitizer and coverage PRs also add §357; whichever merge later renumber.
• CHANGELOG entry under Changed.

🤖 Generated with Claude Code

[jdatcmd]

Reviewed at ee0ade5 by running the guard set, the new guard with the workflow change reverted, a trial merge with main, and the CI query.

Blocking

The branch conflicts with main, so no CI has run. #2036 merged as 1bc280a and took the same registry lines.

$ git merge-tree --write-tree --name-only origin/main ee0ade53
CONFLICT (content): Merge conflict in test/pytest/TESTS.md
CONFLICT (content): Merge conflict in test/pytest/expected_tests.txt
$ gh api repos/commandprompt/pgcolumnar/pulls/2041 -q '.mergeable, .mergeable_state'
false
dirty
$ gh pr checks 2041
no checks reported on the 'ci/2038-aarch64-pytest' branch

To fix, merge main into the branch and:

1. In test/pytest/TESTS.md, renumber this section and its table-of-contents entry to 357. Main's §356 is test_extension_upgrade.py.
2. In test/pytest/expected_tests.txt, set guard_tests 868 and keep main's cluster_tests 5350. I resolved the conflict in a trial merge and collected the NO_CLUSTER files: 868 tests collected. The new guard passes there too: 2 passed.

What I verified at the head

• The guard set (NO_CLUSTER files, --pgc-expect-tests 865) passes: 865 passed, accounting: 2309 pass + 0 fail + 0 unrun.
• Red: with only the nightly.yml hunk reverted, the property arm fails for the reason in #2038:
  AssertionError: pytest-cluster runs every (major, architecture) the suites job runs: got 'nightly.yml:18/aarch64' want 'none' (1 failed, 1 passed).
• PyYAML parses the new pytest-cluster matrix, and it equals the suites matrix exactly (True).
• Every pin in requirements-test.txt has a cp312 manylinux aarch64 wheel on PyPI, or is pure Python (pytest, pytest-xdist).
• The apt packages are the same ones the suites aarch64 leg already installs. That leg passed in last night's run, 42373475655.
• Nothing else uses the old job name pytest (cluster tests, PG N). The reporter's needs: uses the job id, which is unchanged.
• census-commits --against origin/main: 0 commits change the tsv, 0 refused, so no trailer is needed.
• The CHANGELOG, TESTS.md text and the test names agree with the code.

Not blocking

• Timing: in run 42373475655, x86_64 pytest (cluster tests, PG 18) took 31m09s of its 40-minute limit. The aarch64 suites leg was faster than x86_64 that night (12m03s against 14m18s), so the arm pytest leg may fit. That is unverified until the first nightly, as the PR says.
• _arch treats any runner label that contains arm as aarch64. ubuntu-24.04-arm is the only such label today. A narrower test (an -arm suffix) would not misread a future label such as warm-pool.

🤖 Generated with Claude Code

[jdatcmd]

Approving at 8a8ba505. My one blocker on ee0ade53 was the conflict with main after #2036. It's resolved: the branch is now two commits on main 1bc280a1, this PR's TESTS.md section is 357 (main's test_extension_upgrade.py keeps 356), and guard_tests is 868 with cluster_tests 5350, the values my review measured.

Re-checked on this head:

• The nightly.yml matrix carries the five legs: PG 15–18 on x86_64, plus PG 18 on ubuntu-24.04-arm. It's the change my review verified.
• test_pytest_runs_where_the_suites_run.py: 2 passed, 9 pass + 0 fail + 0 unrun.

Nits, still non-blocking:

• _arch treats any label containing arm as aarch64. Matching an -arm suffix would be stricter.
• Whether the aarch64 leg fits its 40-minute limit only shows on the first nightly.

Merge order: my #2039 also uses TESTS.md section 357 at the moment, so whichever of us lands second renumbers.

🤖 Generated with Claude Code