Flaker Storage Cache On CI

flaker storage cache on GitHub Actions

flaker keeps its data in a DuckDB file at the path declared by flaker.toml [storage] path (default .flaker/data). For flaky detection / KPI / quarantine to work, that file must persist between CI runs. GitHub Actions has no first-class runtime storage, so the convention is actions/cache@v4 with a sliding key.

When this skill applies

• "新しい repo に flaker を CI で動かす"
• "flaker の履歴が CI で消えてる / 毎回ゼロからになる"
• "flaker に を import するワークフロー追加して"
• "VRT / playwright / custom-adapter report を flaker に流したい"

Cache key shape

- name: Cache flaker data
  if: always()
  uses: actions/cache@v4
  with:
    path: .flaker/data
    key: flaker-data-${{ github.run_id }}
    restore-keys: |
      flaker-data-

• key uses github.run_id so each run writes a new cache entry on save (GH Actions saves at end of job).
• restore-keys: flaker-data- matches anything written previously, falling back to the most recent — so the next run sees the latest accumulated state regardless of which earlier run produced it.
• if: always() so a failed earlier step still triggers the save (history-on-failure is fine and often more useful than history-only-on-success).

The path MUST equal flaker.toml's [storage] path. Default is .flaker/data (a file, not a directory — pre-creating it as a dir breaks the DuckDB open).

In a pnpm workspace monorepo: prefix the cache path with the package directory, e.g. packages/<pkg>/.flaker/data. flaker does NOT walk up to find flaker.toml — it must live next to where flaker is invoked. Mismatched cache path vs flaker.toml location surfaces as Config file not found (when running from repo root) or "history vanished every run" (when cache and flaker disagree on where storage is).

Triggering writes

actions/cache@v4 saves automatically in its post-step. Callers don't cache save explicitly. The save key is the run-id, so duplicate writes never collide.

fetch-depth for --changed derivation

Any flaker invocation that uses the hybrid / affected strategy (flaker run --gate merge in CI profile, by default) needs --changed <files,...>. Without it: Error: hybrid mode requires resolver and changedFiles.

- uses: actions/checkout@v4
  with:
    fetch-depth: 0  # need history for `git diff` against the PR base

- name: flaker run --gate merge
  env:
    BASE_REF: ${{ github.event.pull_request.base.ref }}
  run: |
    changed=$(git diff --name-only "origin/${BASE_REF}...HEAD" | tr '\n' ',' | sed 's/,$//')
    pnpm exec flaker run --gate merge --changed "$changed"

Empty diff (config-only PR, etc.) is fine — hybrid falls back to the configured fallback strategy.

Workflows: which one persists, which one reads

Reference layout for splitting workflows by trigger:

Rule of thumb: anything that touches main / scheduled writes; PR-scoped checks read-only or no cache.

Adding a new ingest source

Common pattern (custom adapter, see flaker#79 for the adapter contract):

1. Produce the report file (e.g. <source>-report.json, vitest-report.json).
2. Restore the cache (or rely on the same job's earlier restore step).
3. Import:

   - name: Import <source> into flaker
     if: always()
     run: |
       pnpm exec flaker import <report-file> \
         --adapter <name> \
         --commit "${{ github.sha }}" \
         --branch "${{ github.ref_name }}" \
         --source ci
   

4. The cache save at end-of-job picks up the new rows automatically.

if: always() ensures partial-failure runs still record what they saw before the failure.

Don't do this

• Don't pre-create .flaker/data as a directory before flaker runs. DuckDB expects to open it as a file; IO Error: Could not read from file ... Is a directory is the symptom. The cache path: .flaker/data referencing a not-yet-existing file is fine — actions/cache restores it if present, and flaker creates it if not.
• Don't add flaker import to PR-only workflows without thinking. PR runs happen on every commit-to-PR and would dominate the population. Keep PR jobs read-only against the cache.
• Don't use NODE_OPTIONS=--preserve-symlinks-main with flaker ≥ 0.10.7. It was a workaround for a pnpm symlink bug fixed in 0.10.7; under 0.11.x the env var silently turns the CLI into a no-op, exit 0 (no output, no work done). Surfaced via downstream usage, removed in flaker upstream.
• Don't try to write the cache from a PR fork. GH Actions disallows writes from forks for security reasons; the restore still works, the save silently no-ops.

Diagnostics

If history "vanishes":

1. Check the cache hit log line in the workflow run — if it says Cache not found for input keys, the restore-keys prefix doesn't match what was saved.
2. Check flaker.toml [storage] path matches the cached path: exactly.
3. Run flaker doctor in the same job — confirms the file is open-able.
4. flaker status --markdown exposes row counts so you can see what was actually loaded.

Reference

• flaker itself (docs/contributing.md) for the storage path convention.
• The --adapter system docs for writing a custom report importer.