Revisions with `initialScale > 1` that are no longer referenced by any Route cannot scale down to 0

[Papilionidae]

Problem Description

Revisions with initialScale > 1 that are no longer referenced by any Route (i.e., routingState = "reserve") cannot scale down to 0, causing resource waste when old revisions are replaced by new ones.

Expected Behavior

When a revision's routingState becomes "reserve" (meaning it is no longer referenced by any Route), it should be able to scale down to 0 immediately, regardless of initialScale configuration. This is important because:

• When a new revision becomes ready and replaces the old one, the old revision's routingState changes to "reserve"
• These old revisions should free up resources quickly by scaling down to 0
• The initialScale constraint should not prevent this cleanup

Actual Behavior

Revisions with routingState = "reserve" and initialScale > 1 cannot scale down to 0 because:

1. The initialScale logic in scaler.go forces min = initialScale even when the revision is no longer referenced by any Route
2. This prevents the autoscaler from scaling down below initialScale, even though ScaleBounds() already returns min=0 for unreachable revisions (which includes routingState = "reserve")

Steps to Reproduce

1. Deploy a Knative Service with minScale=1 and initialScale=2:

apiVersion: serving.knative.dev/v1
   kind: Service
   metadata:
     name: helloworld
   spec:
     template:
       metadata:
         annotations:
           autoscaling.knative.dev/minScale: "1"
           autoscaling.knative.dev/initialScale: "2"
       spec:
         containers:
         - image: gcr.io/knative-samples/helloworld-go

2. Deploy a new revision with an invalid image (e.g., imagePullBackOff):

spec:
     template:
       spec:
         containers:
         - image: invalid-image:latest

3. The new revision will have 2 pods in ImagePullBackOff state and will be marked as Unreachable
4. Deploy a third revision with a valid image
5. Observed: The second revision (with ImagePullBackOff pods) remains at 2 pods and cannot scale down to 0, even though the old revision (helloworld-00002) has routingState = "reserve" (no longer referenced by Route):

$ kubectl get po -n paas-uat
NAME                                                  READY   STATUS             RESTARTS   AGE
helloworld-nodejs-00002-deployment-564896c9fc-v7ntx   0/2     ImagePullBackOff   0          3h
helloworld-nodejs-00002-deployment-564896c9fc-vsqrh   0/2     ImagePullBackOff   0          3h
helloworld-nodejs-00003-deployment-847f88dbd8-6vfll   2/2     Running            0          168m

Root Cause Analysis

The issue is in serving/pkg/reconciler/autoscaling/kpa/scaler.go, in the scale() method:

// Line 343-349
if initialScale > 1 && !pa.Status.IsScaleTargetInitialized() {
    // Ignore initial scale if minScale >= initialScale.
    if min < initialScale {
        logger.Debugf("Adjusting min to meet the initial scale: %d -> %d", min, initialScale)
    }
    min = intMax(initialScale, min)
}

This code forces min = initialScale regardless of whether the revision's routingState is "reserve". However:

• When a revision's routingState = "reserve", it means it's no longer referenced by any Route
• ScaleBounds() already returns min=0 for unreachable revisions (see pa_lifecycle.go:90), and routingState = "reserve" results in Reachability = Unreachable (see revision/resources/pa.go:77-78)
• The initialScale logic should respect the routingState = "reserve" condition and not override the min=0 value

Relationship Between routingState and Reachability

According to serving/pkg/reconciler/revision/resources/pa.go:

• routingState = "active" → Reachability = Reachable
• routingState = "reserve" → Reachability = Unreachable
• routingState = "pending" or unset → Reachability = Unknown

So routingState = "reserve" is equivalent to Reachability = Unreachable for the purpose of determining whether a revision should be allowed to scale down.

Proposed Solution

Modify the initialScale check to ignore initialScale when the revision's routingState = "reserve" (i.e., when pa.Spec.Reachability == ReachabilityUnreachable):

if initialScale > 1 && !pa.Status.IsScaleTargetInitialized() && pa.Spec.Reachability != autoscalingv1alpha1.ReachabilityUnreachable {
    // Ignore initial scale if minScale >= initialScale.
    if min < initialScale {
        logger.Debugf("Adjusting min to meet the initial scale: %d -> %d", min, initialScale)
    }
    min = intMax(initialScale, min)
}

This change ensures that:

1. Revisions with routingState = "reserve" (no longer referenced by Route) can scale down to 0 immediately
2. The initialScale logic only applies to revisions that are still referenced by Routes (routingState = "active")
3. Resources are freed promptly when old revisions are replaced by new ones

Environment

• Knative Serving version: Knative v1.19.6
• Autoscaler configuration: initialScale=2, minScale=1

[aviralgarg05]

I would like to solve this issue

[Papilionidae]

@aviralgarg05
Thank you for taking on this issue! I'm happy to help with any questions or provide more details if needed. Looking forward to your fix!

[linkvt]

@aviralgarg05 nice, you can assign yourself to this ticket by writing a comment containing the line /assign (if you still plan to work on this).

[aviralgarg05]

/assign

[aviralgarg05]

Alright!!