The decaying failure window let a repeating cycle of one failure and a run
of rejections readmit a dead backend once per cycle, forever: any rule that
readmits on "no real failure lately" flaps under a periodic pattern.
Replace it with a per-backend latch: has this probe endpoint ever answered
with a verdict we can read? Until it has, there is no health signal, so the
backend is never gated and stays in service as probe_unsupported. Once it
has, the path works and every unsuccessful probe counts, rejections
included. The latch never clears, so no pattern can argue a backend back in.
A backend whose health_probe_path is wrong rejects every probe, so no
success can ever arrive to clear the sticky run-failed flag. One transient
failure in the middle of those rejections excluded such a backend from the
pool for the life of the process, reintroducing the stranding bug.
Decide unhealthy vs probe_unsupported on whether a real failure landed
within the last health_probe_failures probes, floored at 3 so the window
always spans a reject/reject/failure cycle. Sustained alternation keeps a
failure in every window and stays unhealthy; an aged-out failure leaves a
pure-rejection run on probe_unsupported and back in service.
unkin-agent