Description
- implement a readiness probe for the sidecar
- return "ready" only once connected to the driver
Background: we debugged a pmem-csi deployment case where node-driver-registrar 1.0.2 failed to connect to /csi/csi.sock, getting timeout and causing pod exit in 60 seconds, leading to CrashLoop.
We then tried version 1.1.0 which does not timeout. There the pod remained running and gave false impression of good state, but in reality it kept trying socket connection which never succeeded. In that sense the v1.1.0 "try without timeout" semantics can be considered harmful as problem state remains hidden (if one does not look in log).
@pohly then expressed idea about readiness probe.
The deployment problem looked initially a lot like #36 (but SELinux was not involved) but got explained as missing /csi/csi.sock caused by chain starting with other misconfiguration, so nothing to blame as system functioned exactly the way it was configured.