andrijapanicsb commented on PR #14459:
URL: https://github.com/apache/cloudstack/pull/14459#issuecomment-6893015338

   ## Live test results
   
   **In short:** this PR was tested on a real CloudStack with real Kubernetes 
clusters, through the API, the web UI, and with the real cluster-autoscaler. A 
scale request that names no compute offering now leaves the existing nodes 
alone, and ordinary clusters with one offering keep working as before. All 20 
checks passed.
   
   Tested on packages built from `9614df02ac` (24.0.0-SNAPSHOT, EL9), two 
management servers, two KVM hosts, Kubernetes 1.28.4.
   
   ### Cluster with separate control and worker offerings (API)
   
   "Out of date" means the workers were already on a bigger offering (3 GB) 
while the cluster-wide value still said 2 GB. This is the situation from #14552.
   
   | Situation | What happened | Outcome |
   |---|---|---|
   | Scale out by size only | One worker added, control node kept its own 
offering | PASS |
   | Remove one named worker (#14458) | Only that worker removed | PASS |
   | Autoscaling on, then off | Only the settings changed | PASS |
   | Autoscaling on plus new worker offering in one request | Workers resized, 
autoscaling stayed on with min and max kept | PASS |
   | Out of date: autoscaling off | No node resized | PASS |
   | Out of date: scale out by size only | New worker on 3 GB, others untouched 
| PASS |
   | Out of date: remove one named worker | Only that worker removed | PASS |
   | Out of date: real cluster-autoscaler adds a node under load | New worker 
on 3 GB, others untouched, waiting pods started | PASS |
   | Out of date: real cluster-autoscaler removes an idle node | Empty worker 
removed, others untouched | PASS |
   | Explicit cluster-wide offering (behaviour that must stay) | Every node 
resized to it | PASS |
   
   ### Plain cluster, one offering for all nodes (API)
   
   | Situation | What happened | Outcome |
   |---|---|---|
   | Scale out by size only | New worker on the cluster's one offering | PASS |
   | Scale in by size only | One worker removed, nothing resized | PASS |
   | Remove one named worker | Only that worker removed | PASS |
   | Autoscaling on, then off | Settings saved and cleared, nothing resized | 
PASS |
   | Explicit cluster-wide offering change | Both nodes resized, cluster-wide 
value updated | PASS |
   
   ### Web UI
   
   | Action | What happened | Outcome |
   |---|---|---|
   | Scale dialog: autoscaling off | No node resized | PASS |
   | Scale dialog: size 2 to 3 | One worker added, others untouched | PASS |
   | Scale dialog: autoscaling on plus worker offering change, one submit | 
Workers resized, control untouched, autoscaling kept | PASS |
   | Scale dialog: autoscaling off again | No node resized | PASS |
   | Instances tab: Delete Node | Only that worker removed | PASS |
   
   The UI run also showed which callers were exposed. The Scale dialog always 
sends the current control and worker offerings, so it never hit #14552. The 
Delete Node button sends only the node ID, which is the #14458 case. API 
clients and the cluster-autoscaler send no offering at all.
   
   ### Not covered
   
   - The failing behaviour was not re-run on a build without this PR. It was 
seen live earlier, during testing of #14223: see #14458 and #14552.
   - Size-only scale-out with an out-of-date offering, and the autoscaler's own 
scale-up, were never seen failing live. They were tested with the fix only.
   - etcd node groups, VPC networks and stopped clusters were left out on 
purpose. The PR decides from the request alone, before network type or cluster 
state is looked at, and etcd is skipped by the same condition as the control 
group.
   
   ### Noticed in existing code, not caused by this PR
   
   - Cluster CPU and memory totals stay wrong after an offering change. #14223 
rewrites that calculation.
   - An explicit cluster-wide request on a cluster with separate offerings 
resized every node but did not update the cluster-wide value. Same subject as 
#14702.
   
   Full report with screenshots: attached PDF.
   


-- 
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]

Reply via email to