acoustic-carpenter-78188
05/05/2026, 9:53 PMautoscaling.knative.dev/max-scale: "0" as a stop signal, but in Knative that value represents “unlimited” upper bound, so it didn’t reliably enforce the intended stop semantics. Separately, the UI polls/Get calls can briefly show stale “stopped” state during stop→start transitions because the control plane can cache transitional responses while the data plane is still converging.
What changes were proposed in this pull request?
• Replaced the incorrect max-scale=0 stop semantics with Knative-native behavior:
• Stop() labels the Knative Service networking.knative.dev/visibility=cluster-local so it is not published to the external gateway (public ingress becomes inaccessible).
• Stop() marks the Service with flyte.org/app-stopped=true as the control-plane source of truth for STOPPED state.
• Stop() sets autoscaling.knative.dev/min-scale="0" and autoscaling.knative.dev/initial-scale="0" to converge cleanly to zero pods.
• Stop() deletes the latest ready Revision so existing pods terminate immediately (no waiting for the stable window).
• Deploy() (“Start”) clears the stopped/private labels so the service becomes externally routable again.
• Status mapping now treats STOPPED based on flyte.org/app-stopped (not max-scale).
Devbox / cluster configuration
• Enabled config-autoscaler.allow-zero-initial-scale="true" in devbox-bundled kustomize overlays and rendered manifests.
• Updated docker/devbox-bundled/Makefile setup-knative to also patch config-autoscaler so the “manual install” path matches the bundled manifests behavior.
UI correctness / control-plane behavior
• Added a small TTL cache in the control-plane AppService for Get responses to reduce cross-plane chatter, while explicitly avoiding caching “transitional” stop/start windows (desired state running, but status still stopped) so the UI doesn’t get stuck showing STOPPED after clicking Start.
How was this patch tested?
• Unit tests:
• go test ./app/internal/k8s
• go test ./app/service
• Manual verification (devbox / cluster):
• Create app → confirm public URL reachable.
• Stop app → confirm pods scale to 0 and public URL is not reachable (service is cluster-local).
• Start app → confirm visibility label cleared, service becomes publicly reachable again, and UI
transitions to ACTIVE without being stuck in STOPPED due to caching.
Setup process
Screenshots
Check all the applicable boxes
• I updated the documentation accordingly.
• All new and existing tests passed.
• All commits are signed-off.
Related PRs
Stack
If you do use git town to manage PR Stacks, the stack relevant to this PR
will show below. Otherwise, you can ignore this section.
Docs link
flyteorg/flyteacoustic-carpenter-78188
05/06/2026, 10:54 PM