decentraland / comms-gatekeeper / 34891158967

87%

main: 86%

Build Type

Pull #283

github

Committed by LautaroPetaccio

Commit Message

feat: drain in-flight work on shutdown instead of dropping it

Components stop in reverse creation order, so the subscriber stops before
the NATS adapter closes its connection. Nothing used that: a mint that was
mid-flight when the signal arrived reached its publish after the connection
had gone, the adapter dropped the write, and that peer waited for its next
event to get a room.

Three parts make the order count. The NATS adapter's subscribe returns a
handle: it drops the registration first, so a reconnect cannot restore it,
then cancels the live subscription when the link is up. The cluster
subscriber cancels its three subscriptions on stop, so later events go to
another replica in the queue group. The keyed queue then drains, re-checking
each round so work queued during the drain is waited on too.

The drain is bounded by KEYED_QUEUE_DRAIN_TIMEOUT_MS (5 s default, 0 to skip
the wait): the lifecycle has no stop timeout of its own, and a task stuck on
a request without one would hold shutdown until the orchestrator killed the
process. Giving up names the keys still running. Work enqueued after stop is
still accepted, since refusing it would gain nothing.

Pull Request Pull Request #283: feat: mint LiveKit island rooms from Pulse cluster feed (Archipelago ⇒ Pulse, iteration 1)

Coverage Stats

1265 of 1575 branches covered (80.32%)

Branch coverage included in aggregate %.

354 of 357 new or added lines in 17 files covered. (99.16%)

4 existing lines in 2 files now uncovered.

3475 of 3866 relevant lines covered (89.89%)

82.7 hits per line