Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Manually scale Pod Autoscaler of a revision #15480

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Manually scale Pod Autoscaler of a revision #15480

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 5 additions & 3 deletions pkg/reconciler/revision/reconcile_resources.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,12 +173,14 @@ func (c *Reconciler) reconcilePA(ctx context.Context, rev *v1.Revision) error {
// We no longer require immutability, so need to reconcile PA each time.
tmpl := resources.MakePA(rev, deployment)
logger.Debugf("Desired PASpec: %#v", tmpl.Spec)
if !equality.Semantic.DeepEqual(tmpl.Spec, pa.Spec) {
diff, _ := kmp.SafeDiff(tmpl.Spec, pa.Spec) // Can't realistically fail on PASpec.
logger.Infof("PA %s needs reconciliation, diff(-want,+got):\n%s", pa.Name, diff)
if !equality.Semantic.DeepEqual(tmpl.Spec, pa.Spec) || !equality.Semantic.DeepEqual(tmpl.Annotations, pa.Annotations) {
Copy link
Contributor

@skonto skonto Sep 25, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @saileshd1402 I suppose you want to avoid creating a new revision, what is the use case you have? Could you describe the problem that you have?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @skonto, essentially we are facing issues with manually scaling/updating replicaset. By default when we change scale, knative-serving brings up the pods associated with the new revision, wait for them to be active then scales down the previous revision. This doesn't work for on prem scenarios since there are fixed limited resources especially when dealing with GPUs so the new revision never gets ready.

Copy link
Contributor

@skonto skonto Oct 3, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Hi @saileshd1402 this is something https://github.com/knative-extensions/serving-progressive-rollout tries to address. You can ask @houshengbo on this one, they are facing the same in their org. Vincent maintains the extension and has some ideas on the topic.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We did try using progressive rollout extension but it had it's own issues we faced:

  1. Can there be an approach to terminate the last pod of the previous revision when most pods of new revision is up knative-extensions/serving-progressive-rollout#200
  2. With the resourceUtil strategy, graceful traffic transfer does not occur during consecutive update requests. knative-extensions/serving-progressive-rollout#202
  3. With the resourceUtil strategy, traffic transfer does not occur during an update when deployment hits quota limit knative-extensions/serving-progressive-rollout#203

In a brief summary, we have noticed that the "resourceUtil" strategy fails to do graceful traffic transfer during consecutive updates, particularly when resource limits are hit. This leads to stuck states and failed requests, as traffic may still continue to direct to terminated revisions while new pods remain in pending state.

diffSpec, _ := kmp.SafeDiff(tmpl.Spec, pa.Spec) // Can't realistically fail on PASpec.
diffAnnotations, _ := kmp.SafeDiff(tmpl.Annotations, pa.Annotations)
logger.Infof("PA %s needs reconciliation, diff(-want,+got):\n%s\n%s", pa.Name, diffAnnotations, diffSpec)

want := pa.DeepCopy()
want.Spec = tmpl.Spec
want.Annotations = tmpl.Annotations
if pa, err = c.client.AutoscalingV1alpha1().PodAutoscalers(ns).Update(ctx, want, metav1.UpdateOptions{}); err != nil {
return fmt.Errorf("failed to update PA %q: %w", paName, err)
}
Expand Down