LivePulse

Test failed

A Playwright E2E test published a Redis update once the client's WebSocket reached readyState OPEN, then asserted the new score appeared. It passed reliably on a local machine but failed consistently in CI — the page kept showing the original fixture score.

Initial assumption

The failure looked like a timing/flakiness issue, so the first attempt bumped the assertion timeout to 10 seconds. It didn't help — nothing was slow, something was simply lost.

Investigation

The CI trace's actual page snapshot showed the pre-update score still on screen, ruling out 'just needs a longer timeout' before that fix was even tried.

Root cause

A client socket reaching readyState OPEN only proves the WebSocket handshake finished — it says nothing about whether the server has issued its Redis SUBSCRIBE for that match's channel yet, which happens asynchronously after the server receives the client's subscribe message. The test published to Redis in exactly that window. Redis pub/sub has no delivery guarantee for a message published before a subscriber exists (a documented ADR-006 tradeoff) — CI's colder first database query widened the race window enough to lose it consistently, where it was apparently always won locally.

Fix

Changed the test to wait for the actual match:snapshot message — which the server sends only after it has subscribed to Redis — instead of waiting on the client socket's readyState. No application code changed; the gateway's subscribe-then-snapshot ordering was already correct, only the test's synchronization assumption was wrong.

Verification

6/6 runs passed locally across 3 repeated executions against a freshly seeded, isolated backend mirroring CI's own setup, each completing in under a second — where the old race-prone version needed the full retry-and-timeout path to occasionally pass at all.