younsl commented on PR #386:
URL: 
https://github.com/apache/superset-kubernetes-operator/pull/386#issuecomment-5825236074

   @villebro Thanks, agreed on all three. Addressed in 1704564, with 2 split 
out as you suggested.
   
   **1. Budgets that block every voluntary eviction.** You were right that 
"never blocks node drains" only held for the unset fallback. Rendering now 
fails for an integer minAvailable greater than or equal to replicas and for 
maxUnavailable 0, with the failedTemplate cases you asked for plus the passing 
replicas 2 and minAvailable 1 case. I went one step further on percentages: 
since this chart has no HPA, replicas is static, and the disruption controller 
rounds both bounds up against it, percentages turn out to be exactly checkable 
at render time. So maxUnavailable 0% and any minAvailable percentage that 
rounds up to every replica (for example 67% at three replicas, while 66% is 
accepted) are rejected too. I confirmed the rounding model against 
status.disruptionsAllowed on a kind cluster for six configurations. The docs 
still tell users to check that percentage bounds leave at least one disruption.
   
   **2. replicas > 1 without leader election.** Split into #390 to keep this PR 
focused. It fails rendering in deployment.yaml as you proposed and adds a 
changelog entry under Changed, since existing releases with that combination 
will now fail helm upgrade.
   
   **3. HA docs.** Added a High availability section to the installation guide 
covering active/standby with leader election required, roughly 15 second 
failover because LeaderElectionReleaseOnCancel is off, spreading replicas 
across nodes, and enabling the PDB only with two or more replicas. The 
podDisruptionBudget.enabled description now says "Only useful with replicas >= 
2", and the changelog entry is corrected.
   
   One possible follow-up: enabling LeaderElectionReleaseOnCancel would make 
the leader release the Lease when a drain evicts it, so the standby takes over 
almost immediately instead of after about 15 seconds. That fits well with the 
PDB, but the scaffold comment in cmd/main.go warns it is only safe if the 
process exits right after the manager stops, so I would like to check the 
shutdown path first. Happy to open a separate PR if you think it is worth it.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to