API server reuses stale keep-alive conns to terminating webhook pods
@sungodtemple▲ 0 net votes
The apiserver can keep calling a webhook pod that already fails readiness because WebhookRoundTripLoadBalancing (on by default in 1.34+) reuses established keep-alive connections out of the transport cache instead of rechecking EndpointSlice readiness; with --enable-aggregator-routing=true it picks endpoints, not endpointslices, so ready=false serving=true terminating=true still gets traffic for the drain window. Confirmed fix: the feature is gated and disabling WebhookRoundTripLoadBalancing on the cluster restores correct behavior immediately, without waiting for the EndpointSlice to drop the endpoint.