Only retry what can succeed
Pass the HTTP statusCode on each attempt. A failure whose status is
not in 408, 425, 429, 500, 502, 503, 504 fails the same way on every attempt.
A 400, 401, 403, 404 or 422 won't change on its own, so retrying it only burns
budget. retryfuse returns brake with evidence
fail and reason
non_retryable_status on the first such failure, without waiting for
the failure streak.
Full-jitter exponential backoff
When the decision is replay_safe after a
failure, the report carries retryAfterMs: a random delay between 0
and min(maxDelayMs, baseDelayMs × 2^streak). Spreading callers across
the whole window stops them retrying in lock-step and hitting a recovering
service all at once. Defaults are 100ms
base and 30000ms cap
(policy.baseDelayMs, policy.maxDelayMs). The jitter is
seeded from the operation rather than a random source. The same input always
gives the same delay, so CI runs are reproducible.
Retry-After is a floor
If the server sent a Retry-After (pass it as the attempt's
retryAfterMs), the returned delay is never shorter than that hint.
Jitter can only lengthen it, and maxDelayMs still caps it.
Half-open after a cooldown
Set policy.cooldownMs, pass now, and stamp each attempt
with at. A circuit tripped by the failure streak stays
brake until the cooldown has elapsed since
the last failure. After that it goes half-open and admits one probe
(circuit_half_open). If the probe fails, the circuit re-opens; if it
succeeds, the circuit closes. The budget brake still applies. The engine never
reads a clock; you pass now in.
Correctness still comes first
None of this applies to an uncertain write. A post-write timeout returns
reconcile_required with evidence
unknown whatever its status code or
backoff. Waiting longer does not make a possibly-applied effect safe to replay.
See why reconcile beats a
blind retry and braking a
retry storm.