A single cluster node with degraded networking (intermittent packet loss) doesn't just get marked unreachable. Its own failure detector gossips false unreachable verdicts against the healthy peers it monitors.
Because Akka currently treats a node as unreachable cluster-wide if any single observer reports it so (no quorum on the verdict), SBR's indirectly-connected downing logic then removes those falsely-accused healthy peers.
Idea: introduce a quorum threshold k for unreachability verdicts: a subject is only considered unreachable cluster-wide when more than k distinct observers independently report it as unreachable, instead of trusting any single observer.
Expected to be opt-in (existing single-observer behavior remains the default).
A single cluster node with degraded networking (intermittent packet loss) doesn't just get marked unreachable. Its own failure detector gossips false unreachable verdicts against the healthy peers it monitors.
Because Akka currently treats a node as unreachable cluster-wide if any single observer reports it so (no quorum on the verdict), SBR's indirectly-connected downing logic then removes those falsely-accused healthy peers.
Idea: introduce a quorum threshold
kfor unreachability verdicts: a subject is only considered unreachable cluster-wide when more than k distinct observers independently report it as unreachable, instead of trusting any single observer.Expected to be opt-in (existing single-observer behavior remains the default).