prisma-risk / tsoracle / 26303181913

95%

Build Type

push

github

Committed by web-flow

Commit Message

fix(client): honor LeaderHint.leader_epoch and TTL the cached leader (#126)

* fix(client): honor LeaderHint.leader_epoch and TTL the cached leader

Track epoch and last-used Instant alongside the cached leader endpoint
so a delayed NOT_LEADER from an old epoch cannot flap the cache
backward, and a cached leader that has fallen quiet past
RetryPolicy::leader_ttl (default 30s) is re-evaluated rather than
pinned indefinitely.

- Replace ChannelPool.leader: Option<String> with
  Option<CachedLeader { endpoint, epoch: Option<u64>, last_used }>.
- accept_hint(hint_epoch) is the gate: reject only when the cache is
  fresh, both epochs are known, and the hint's epoch is strictly less
  than the cached one (monotone-forward).
- record_success(endpoint, epoch) refreshes last_used on hits and
  installs a fresh entry on misses; set_leader_with carries the hint
  path (epoch may be None for old servers).
- RetryPolicy gains leader_ttl: Duration, default 30s (~3x the default
  overall_deadline) so a steady-state leader is retained indefinitely
  but a quiet one re-evaluates the configured list.
- Stale hints are dropped silently and counted as
  tsoracle.client.leader_hint.stale.total.

Hints that omit leader_epoch (current server behaviour) and hints
arriving before the cache has seen any successful RPC are accepted
unconditionally so deployments mid-upgrade are not left without leader
discovery. Server-side population of leader_epoch is tracked in #125.

Refs #87

* test(client): cover classify_not_leader_hint and trim assert noise

Extract the post-FAILED_PRECONDITION decision (decode hint, reject
plaintext under TLS, gate on cached epoch) out of `attempt` into a
pure `classify_not_leader_hint(pool, endpoint, status) -> AttemptOutcome`
helper. Same shape as the existing `accept_hint`,
`rejects_plaintext_hint`, and `should_backoff` predicates — the
production path goes through the helper too, so the integration and
unit tests exercise the same code.

Adds six d... (continued)

Coverage Stats

223 of 239 new or added lines in 4 files covered. (93.31%)

2 existing lines in 1 file now uncovered.

7355 of 7670 relevant lines covered (95.89%)

750826.53 hits per line