prisma-risk / tsoracle / 26414752906

95%

push

github

fix(client): keep the election signal sticky so a final timeout can't bury NOT_LEADER (#432)

issue_rpc recorded every outcome into a single last_err slot, so a
ride-out that recorded a reachable peer's NoLeaderYet (FAILED_PRECONDITION)
could have it clobbered by a DeadlineExceeded from the final attempt —
whose per-attempt budget the overall deadline had squeezed to near zero.
The client then surfaced a transport timeout instead of the cluster's own
'no leader yet' diagnosis, which is both less actionable for callers and
made exhausted_ride_out_surfaces_not_leader flaky on slow CI runners.

Track the in-progress-election signal separately and make it sticky across
passes, then surface errors by precedence in a new pure surface_error():
a non-transport last_err (a definitive server rejection) wins verbatim;
otherwise the sticky election signal; otherwise the transport-class
last_err; otherwise NoReachableEndpoints. A reachable server's NOT_LEADER
status thus outranks a transport straggler that is only an artifact of our
own budget.

Add deterministic unit tests for surface_error (no sockets or timing) and
a readiness warm-up to the integration test so at least one attempt reaches
the peer before the ride-out is timed.

53 of 65 new or added lines in 1 file covered. (81.54%)

10403 of 10781 relevant lines covered (96.49%)

467616.52 hits per line