dunglas / mercure / 24781496320

84%

master: 93%

Build Type

push

github

Committed by web-flow

Commit Message

fix(bolt): stop leaking post-subscribe events into Last-Event-ID header (#1215)

dispatchHistory used to update responseLastEventID with each key the
cursor visited BEFORE checking whether that key was already past the
subscriber's sequence snapshot. Under contention — the subscriber is
added to the transport's subscriber list *before* dispatchHistory opens
its read transaction, so a Dispatch can slip in between — the scan would
observe keys that didn't exist at subscribe time and report one of them
as the "last event" in the response header. Symptom: flaky
TestUnknownLastEventIDEmptyHistory on main ("expected: earliest,
actual: b" whenever the test's live dispatch landed before the read tx).

Two small changes:

- Reorder the loop so pastSeqBound is checked BEFORE responseLastEventID
  is assigned. Keys past the bound no longer touch the response.
- pastSeqBound now uses a strict > comparison (and drops the toSeq > 0
  short-circuit). The previous >= was off-by-one — it broke the loop on
  the exact key at toSeq, i.e. the most recent event observed at
  subscribe time. The short-circuit also meant toSeq == 0 (empty bucket
  at subscribe time) let every subsequent key through. Strict > on a
  bigendian seq compares correctly for all cases including toSeq == 0.

Existing TestUnknownLastEventID still passes (event "a" persisted BEFORE
subscribe keeps showing up as the last event); TestUnknownLastEventIDEmptyHistory
is now deterministic.

Coverage Stats

8 of 9 new or added lines in 1 file covered. (88.89%)

4 existing lines in 1 file now uncovered.

1744 of 2079 relevant lines covered (83.89%)

50.04 hits per line