prisma-risk / tsoracle / 26340894812

96%

push

github

fix(paxos-toolkit): decouple PaxosRunner tick loop from outbound send completion (#218)

A never-resolving `MessageSink::send` wedged the entire tick task. The
serial `for message in outgoing { sink.send(message).await }` ran inside
the `tokio::select!` tick arm, and `select!` does not re-poll its branches
while an arm body is suspended mid-await. One hung send therefore starved
later sends in the batch, leadership observation, leader-event emission,
`apply_notify`, the next tick, and the shutdown branch — so `stop()` blocked
forever awaiting the wedged `JoinHandle`. A blackholed peer reached over a
transport with no per-request timeout (as in the example `PeerSink`)
reproduces this.

Move outbound delivery to a dedicated sender task that owns the sink and
drains a bounded channel. The tick loop hands messages off with a
non-blocking `try_send` and never awaits delivery, so its cadence — and
every lifecycle action downstream of it — is independent of per-send
latency. A full queue drops the message, which is safe because OmniPaxos
regenerates outbound messages every tick. `stop()` and `Drop` abort the
sender task as the backstop for a send that never resolves.

Closes #215

56 of 65 new or added lines in 1 file covered. (86.15%)

10074 of 10627 relevant lines covered (94.8%)

543814.81 hits per line