ci: cancel in-progress E2E matrix runs on the same PR branch

[travagliad]

Why

Every push to a PR branch triggers the full 30-job E2E tests matrix (~30 min). Rapid iteration (e.g. dependency fix attempts) spawns parallel matrix runs with no cancellation, wasting CI minutes.

How

Add a concurrency group to e2e-tests-matrix.yml keyed by the target branch name. cancel-in-progress is enabled only for pull_request events, so:

• PR branches: new push cancels the in-progress matrix run
• main (schedule / manual dispatch): runs are not cancelled — RC testing and nightly jobs can overlap as before
• Different branches: run in parallel (separate concurrency groups)

[travagliad]

Test Healer — PR #1091 validation results

Run: E2E tests Matrix #29782815602
Branch: cursor/workflow-concurrency
Event: pull_request

CI summary (30 matrix jobs)

Result	Count
✅ Pass	30
❌ Fail	0
⏳ Pending	0

All green. The concurrency change does not break the E2E matrix workflow.

---

Change under test

concurrency:
  group: e2e-tests-${{ github.event_name == 'pull_request' && github.head_ref || (github.event.inputs.pmm_qa_branch || github.ref_name) }}
  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

Scenario	Behaviour
PR push to same branch	Cancels in-progress matrix run (new group key = head_ref)
workflow_dispatch on main	No auto-cancel (cancel-in-progress: false)
Nightly schedule on main	No auto-cancel
Different PR branches	Run in parallel (separate group keys)

---

Cross-check vs #1092 (same CI run window)

Jobs that failed on #1092 passed here on unmodified main deps:

Job	#1092	#1091
@pgsm-pmm-integration	❌ Docker network flake	✅ pass (28m55s)
@mongodb-exporter	❌ PMM-T1333 dashboard timing	✅ pass (41m59s)

Confirms those #1092 failures were flakes, not caused by dependency changes.

---

Recommendation

✅ Ready to merge. Low-risk CI-only change; full matrix validated green.

[travagliad]

Concurrency test results

Tested by pushing two commits in quick succession to cursor/workflow-concurrency while the matrix was already running.

Test procedure

1. Push 1 (d34734a) — empty commit at 23:47:47 UTC → run #29788349174 started (in_progress)
2. Push 2 (e4b16b1) — empty commit at 23:48:10 UTC (23s later) → run #29788368254 started
3. Cleanup — force-pushed back to cdb4ab94 → run #29788409249 started, cancelling run 2 as well

Results

Run	Commit	Trigger	Outcome
#29788349174	d34734a	push 1	cancelled after 42s (all started jobs cancelled)
#29788368254	e4b16b1	push 2	cancelled when force-push arrived
#29788409249	cdb4ab94	cleanup	in progress (sole active run)

At no point were two matrix runs active on the same branch — each new push cancelled the previous one within ~20s.

Verified behaviour

• cancel-in-progress: true on pull_request events works as intended
• Concurrency group e2e-tests-cursor/workflow-concurrency correctly deduplicates runs on the same branch
• Probe commits were removed (branch reset to cdb4ab94); only the concurrency YAML change remains

✅ Concurrency feature confirmed working.