Autoscaling: 1->0 deadzone

[akyyy]

/area autoscale
/area networking
/kind bug

Expected Behavior

1->0 should be smooth.

Actual Behavior

There is a deadzone.

[mattmoor]

Is this tracking the race condition?

[akyyy]

The bug is when we're deleting k8s resources, the pod is stuck in "terminating" state and gets stuck. We're seeing 503's during tearing down if traffic comes in.
The pull request is here:
#1251