prisma-risk / tsoracle / 26385673710

95%

Build Type

push

github

Committed by web-flow

Commit Message

fix(client): floor the failed-attempt budget at the worklist size (#404)

`iter_round_robin` starts the configured-endpoint tail at a randomly
seeded per-pool rotation offset so a fleet cold start spreads across
nodes. The retry loop, however, stopped after `max_attempts` failed
attempts. In a pool with more configured endpoints than `max_attempts`
(default 5), an unlucky offset could park the only reachable endpoint
behind more failing peers than the budget allowed, so the loop exhausted
its attempts and returned an error without ever dialing the live peer —
a reachable cluster reported unreachable purely because of where the
random offset landed. The previous deterministic order let operators
pin a known-live endpoint first; rotation silently removed that.

Floor the failed-attempt budget at the initial worklist length, so a
single cold-cache sweep always dials every endpoint it knows about at
least once. Leader-hint redirects still grow the worklist beyond that
and stay uncharged against the budget, and the overall deadline still
bounds the worst case.

`max_attempts` can no longer cut a cold sweep short, so the existing
`max_attempts_caps_iteration` test (whose premise became false) is
reworked into `failed_attempt_budget_is_floored_to_a_full_sweep`, and a
new `rotation_offset_cannot_strand_the_only_live_endpoint` regression
test pins the behaviour.

Coverage Stats

90 of 93 new or added lines in 1 file covered. (96.77%)

1 existing line in 1 file now uncovered.

9309 of 9733 relevant lines covered (95.64%)

660108.14 hits per line