prisma-risk / tsoracle / 26382464574

95%

push

github

fix(client): rotate iter_round_robin worklist over configured endpoints (#395)

* fix(client): rotate iter_round_robin worklist over configured endpoints

iter_round_robin returned [cached_leader?] + configured-in-order on every
call, so the name promised a rotation it never performed and, in a
leaderless window, every pool opened its cold probes on configured[0].

Advance a per-pool cursor per worklist and start the configured tail at
cursor % len, keeping the fresh leader first and deduped. The cursor is
seeded randomly per pool so a fleet cold start spreads across nodes rather
than colliding on the first endpoint; a hash of the endpoint list would
not, since every client shares the same list.

Closes #342

* test(client): make connect-failure metric scenario rotation-agnostic

iter_round_robin now starts the configured tail at a random rotation
offset, so the client-metrics end-to-end test's [dead, live] ordering
assumption became a coin-flip: when the cursor lands on the live peer
the request succeeds on the first dial and retries{connect_failure} /
connect.failures never fire.

Drive the connect-failure signals from a single unreachable endpoint
instead. They fire on the failed dial regardless of rotation offset or
any follow-up endpoint, so the scenario is deterministic and no longer
re-encodes the configured-order guarantee the rotation removed. The
fall-through-and-recover path stays covered by the retry-loop unit
tests.

45 of 45 new or added lines in 1 file covered. (100.0%)

12788 of 13481 relevant lines covered (94.86%)

423517.71 hits per line