prisma-risk / tsoracle / 35763332338

97%

Build Type

push

github

Committed by web-flow

Commit Message

fix(client): refresh the cached leader after a successful lease RPC (#717)

The leader cache stays fresh for leader_ttl after its last refresh, and only the GetTs and GetSeq success paths refreshed it. control_plane_rpc seated a NOT_LEADER hint but ignored its own successes. A caller that only issues lease RPCs therefore kept a seated hint for exactly leader_ttl, however many renewals succeeded against it. The next call fell back to the first configured endpoint. When that endpoint is a follower, or a load-balanced address whose connection is pinned to one, the call was refused again: one failed control-plane call every leader_ttl, forever.

The map closure of each control-plane RPC now returns a ControlPlaneReply: the decoded output plus the epoch at which the reply proves the endpoint led, if it proves that at all. On success control_plane_rpc passes that epoch to ChannelPool::record_success, so the refresh follows the same monotone-forward rule as GetTs. AcquireLease and RenewLease are leader-only and carry an epoch, so they refresh. GetSafeFrontier and GetCurrentMaxSafe do not: followers answer them with a zero value and Epoch::ZERO, which a single-node leader also uses, so a success says nothing about leadership. ReleaseLease does not either: it is leader-only but its reply has no epoch, and an epoch-less confirmed success on an empty cache would seat an entry that a later success from the real leader could not displace.

Tests: lease_successes_keep_the_redirected_leader_cached_past_leader_ttl boots a follower and a leader, points a client with a 300 ms leader_ttl at the follower, follows the redirect, and renews every 75 ms for three TTLs. Without the fix renewal 3 is refused by the follower at about 310 ms. Unit tests cover a renewal seating the serving endpoint at its epoch, a late renewal at a lower epoch not unseating a fresher leader, and release, safe-frontier and max-safe successes leaving the cache untouched. The client usage doc now lis... (continued)

Coverage Stats

145 of 151 new or added lines in 2 files covered. (96.03%)

1 existing line in 1 file now uncovered.

19168 of 19838 relevant lines covered (96.62%)

259682.96 hits per line