Repository navigation
|
Hi, yesterday we have upgraded from operator 1.x to 2.0.3. Previously a webhook patched the pods and inserted a TopologySpreadConstraint. Since 2.0.1 supports this natively we have dropped this workaround and add the TSC to our postgresqls resources. The problem is, that on some instances the operator had a problem reaching Patroni (and everything were fine until we restartet the operator with the new image). Then the operator postponed the pod restarts, so the 3 replicas hat no TSC and were not evenly distributed across dc's. Then after 30 minutes during the regular SYNC it catched up and finally restarted the pods. Here's the relevant log: My question is, can we do something to fix this behavior? Thanks in advance, |
Replies: 1 comment
|
The 30-minute delay matches the documented default In v2.0.3's reconciliation code, I would first verify that the operator pod's network context can reach each database pod on TCP 8008, If the remaining concern is recovery latency after connectivity returns, the documented I would preserve the Patroni safety check rather than force-delete running database pods to obtain the |
The 30-minute delay matches the documented default
resync_period; the postponement itself is a safetycheck, rather than evidence that topology spread constraints are unsupported.
In v2.0.3's reconciliation code,
Patroni configuration/restart errors prevent pending pod recreation when all pods are still running.
Your
no route to hosterrors explain why that branch was taken. They do not establish whether theunderlying cause is CNI routing, a node firewall, NetworkPolicy, or another connectivity problem.
I would first verify that the operator pod's network context can reach each database pod on TCP 8008,
including both
/configand/patroni, and compare a failing pod/node with a reachable …