We have a plugin for Elasticsearch to cluster based on looking up endpoints
on its clustering service (which runs at separate port 9300 instead of http
port 9200). But in order to be among the endpoints on a service, the
cluster members have to be considered "up"; so this must occur before they
can even discover each other. The result is that there can't be a
meaningful readiness probe, and clients of the service get back errors
until it is really up.

We could get around this if readiness probes could be honored/ignored by
specific services, or if there were some other method of indicating a more
nuanced "readiness". If the service for port 9300 could consider the
members up once in "Running" state, but the service at port 9200 waited for
a readiness check, everything would work out well.

Is this strictly a kubernetes issue? Is there any movement in this
direction? It seems like something that many clustered services would
benefit from.
dev mailing list

Reply via email to