Shakedown Flaky Test

Shakedown a flaky test

Goal: turn a flaky test (or a subset of flaky tests) green, deterministically, without ever opening a PR or touching main.

Hard rules — never violate

• NEVER open a pull request on your own.
• NEVER commit to or push main. All work happens on a new branch.
• Before opening or updating any PR, ask the user to explicitly confirm they fully reviewed, understood, and approved everything. (Maintainers treat this as non-negotiable.)

Inputs you need from the user

1. The failing test selector(s) — e.g. distributed/tests/test_client.py::test_client_submit, a file, a -k pattern, or a marker. If the user only gave a test name, expand it to a runnable pytest node id.
2. Examples of failed CI runs (URLs or run IDs). If the user did not provide any, ask for them before doing anything else — the failure signature is what tells you whether a local repro is even plausibly the same bug. Accept:
   • GitHub Actions run URLs (https://github.com/<owner>/<repo>/actions/runs/<run-id>[/job/<job-id>])
   • bare run IDs
   • pasted log snippets Pull the logs with gh run view <run-id> --log [--job <job-id>]. Do not fall back to web_fetch for github.com. If gh does not work, ensure that gh auth status returns all green and if not stop suggest to the user how to fix the issue.

Step 1 — Confirm the test is not deterministic

Run the test once, locally, the normal way:

pixi run test --runslow <test-selector>

If it fails deterministically, stop — this is not a flaky test. Tell the user it fails every time and fix it as a normal bug. If it passes, keep going.

Note down the wall-clock runtime of the run (parse it from the printed durations); you need it for the next step.

Step 2 — Calibrate --count for a ~5 minute local run

pytest-repeat is already a project dependency (pixi.toml). Drive it with --count N:

pixi run test --runslow <test-selector> --count <N>

Compute N from the single-run runtime you noted in Step 1 so the total finishes in ~5 minutes (300s):

N = max(100, round(300 / single_run_seconds))

Rules:

• For a single test node id, N must be at least 100 even if one run is slow (e.g. 10s → still --count 100, accept the long runtime).
• For a subset of tests (file / -k / marker), the 100-minimum applies to the whole selector, not per test — but still aim for ~5 min total. If 100 runs of the whole subset would blow way past 5 min, reduce N so the batch fits ~5 min and say so out loud.

Sanity-check your math by running once with a small --count (e.g. 5) and extrapolating if Step 1's timing was noisy.

Step 3 — Try to reproduce locally

Run the calibrated batch:

pixi run test <test-selector> --count <N>

• If it fails at least once: local reproduction achieved. Go to Step 4.
• If it passes (all N green): no local reproduction. Go to Step 5.

Step 4 — Local reproduction: investigate, fix, verify

1. Investigate. Read the failure from the batch output (and the CI logs the user gave you). Form a hypothesis. Investigate the relevant source code under distributed/. Look at utils_test.py fixtures, timing-sensitive gen_cluster quirks, etc.

2. Produce a fix. Make the smallest correct change. Follow repo style: Black, 120 cols, Ruff rules B/TID/I/UP/RUF.

3. Re-verify with 5 batch rounds. Run the calibrated --count batch 5 separate times (fresh pixi run process each time, so state resets between rounds), or stop early the moment a round fails:

   for i in 1 2 3 4 5; do
     echo "=== round $i ==="
     pixi run test <test-selector> --count <N> || { echo "FAIL round $i"; break; }
   done
   

   • If a round fails: the fix is wrong or incomplete. Go back to 4.1, revise, re-run the 5 rounds.
   • If all 5 rounds are green: done. Report to the user what you changed and why, ask for explicit review approval before any PR. END.

Step 5 — No local reproduction: drive CI torture

You cannot reproduce locally, so reproduce in CI via the torture-test workflow.

5a. Confirm the torture-test workflow exists

gh workflow list --all | grep -i torture

If torture-test is not present, stop and tell the user. Do not create it yourself unless explicitly asked.

5b. Calibrate --count for ~30 minutes of CI

CI is slower and noisier than local. Compute N for a ~30 minute CI run:

N = max(100, round(1800 / estimated_single_run_seconds))

Use the Step 1 single-run runtime as the estimate, then inflate by a CI overhead factor (CI is typically ~1.5–3x slower for distributed's gen_cluster tests; pick a factor from the first torture run's actual timing and refine). For a single test, N must be at least 100.

5c. First torture run — confirm the failure reproduces in CI

Trigger the workflow on the current branch via workflow_dispatch:

gh workflow run torture-test \
  --ref <current-branch> \
  -f test=<test-selector> \
  -f count=<N>

(Adjust -f input names to match the workflow's actual inputs: definitions — read the workflow file first with gh workflow view torture-test --yaml or read on .github/workflows/<file>.)

Wait for it to finish:

gh run watch <run-id>      # or poll: gh run list --workflow torture-test --limit 1

Then pull the logs and verify the failure happens at least once:

gh run view <run-id> --log | grep -E "FAILED|Error|assert"

• If the failure does not reproduce in CI either: tell the user. Ask whether to bump N / widen the selector / try a different environment. Do not loop forever on a non-reproducing failure.
• If it does reproduce: continue.

5d. Get off main

If the current branch is main (check with git branch --show-current), create and switch to a new branch before writing anything:

git switch -c torture-<short-test-name>

Never commit to main.

5e. Write a tentative fix, commit, push

• Write the smallest plausible fix (you're flying blind without a local repro, so lean on the CI failure signature).

• Commit with a clear message.

• Push to the user's fork (the origin remote must point at the fork; verify with git remote -v):

  git push -u origin HEAD
  

  If origin points at dask/distributed (upstream) and the user has no fork remote, stop and ask the user to add their fork as a remote / tell you the fork's URL. Never force-push to upstream.

5f. Trigger torture-test on the fork

gh workflow run torture-test \
  --ref <current-branch> \
  -f test=<test-selector> \
  -f count=<N> \
  --repo <user-fork-owner>/<repo>

(Use --repo to target the fork if the workflow lives there; if the workflow only exists upstream, trigger upstream's workflow_dispatch with --ref <current-branch> — but only if the branch is pushed to upstream, which it should not be for a fork. Prefer running the workflow from the fork. Confirm with the user which repo owns the workflow if unclear.)

Wait for CI to finish (gh run watch / gh run list), then check the result:

gh run view <run-id> --log | grep -E "FAILED|Error|assert|passed|failed"

5g. Loop until green

Repeat indefinitely:

fix → commit → push → trigger torture-test on fork → wait for CI → check result

• If CI is all green: report the fix and a link to the final green run to the user. Ask for explicit review approval before any PR. END.
• If CI still fails: read the new failure, revise the fix, re-commit, re-push, re-trigger. Keep iterating. Each iteration is a new commit on the same branch.
• At the end of every CI run, re-check whether the failure has changed character; if it has, re-read the logs from scratch rather than fixating on the original hypothesis.

Stopping conditions

• Test fails deterministically (Step 1) → handle as normal bug.
• torture-test workflow missing (Step 5a) → stop, ask user.
• No fork remote and no way to push the branch (Step 5e) → stop, ask user.
• Failure reproduces neither locally nor in CI across a couple of bumped-N attempts (Step 5c) → stop, report, ask user how to proceed.
• All verification rounds green (Step 4 or Step 5g) → report, END.

Notes

• Local test command: pixi run test <selector>. CI-equivalent (with coverage, leaks, slow tests): pixi run test-ci <selector> — use this if you want to match CI behaviour more closely locally.
• --count comes from pytest-repeat (already in pixi.toml). Do not add the dependency.
• CI timeout per test is 300s (signal-based on Linux). Keep individual runs well under that or pytest-timeout will kill them and look like a flake.
• Resource-leak detection (--leaks=fds,processes,threads) is on in test-ci; leaks can masquerade as flakiness — keep them in mind when reading failures.
• Always read the workflow file before dispatching so you use the correct input names and --ref semantics.