The kubernetes/enhancements repository had one commit in the last week. It is a documentation change, but it matters if you run a cloud controller manager. KEP 5237 changes how the route controller decides when to talk to your cloud provider, and this commit fixes how you are told to watch it.
What the commit changes ¶
The change is the commit that aligns KEP 5237 metrics with the implementation, committed on 6 November 2026. It touches two files and nets 8 insertions and 5 deletions. The commit message says the design now matches what landed in kubernetes/kubernetes#140147 and #140824.
Before, the KEP promised two metrics:
route_controller_route_sync_total, not gated.route_controller_route_corrections_total, gated.
After, there is one. The corrections counter is removed from the kep.yaml metrics list. Its job moves into labels on the sync counter.
The new label scheme ¶
route_controller_route_sync_total now carries two labels.
trigger:node_changewhen a relevant Node update caused the reconcile,periodicwhen the resync timer did. With the feature gate off, every reconcile counts asperiodic.outcome:errorif the reconcile failed,changedif at least one route was created or deleted, otherwisenoop.
The old corrections metric is now a query. With the gate on, trigger="periodic",outcome="changed" counts routes that a periodic resync had to fix. That is the signal the KEP wants: if periodic passes keep finding work, the node update filters are missing events.
A rough PromQL shape, assuming you scrape the CCM:
sum(rate(route_controller_route_sync_total{trigger="periodic",outcome="changed"}[15m]))
One tradeoff is worth naming. A single counter with labels is cheaper to document and keeps the A/B comparison (gate on versus off) on one series family. The cost is that you can no longer tell the two states apart by metric name. You need the label selector in every alert.
Why the route controller is being changed at all ¶
The README states the motivation plainly. The route controller reconciles on a fixed interval, 10s by default. That produces API requests to the infrastructure provider whether or not anything changed. The service and node controllers already use watches. KEP 5237 moves the route controller to the same model, behind the CloudControllerManagerWatchBasedRoutesReconciliation gate.
The kep.yaml still lists the gate as alpha with a v1.35 alpha milestone, v1.37 beta, and v1.38 stable. The latest milestone is v1.36. The commit does not change any of those dates.
Rollback guidance that came with it ¶
The commit also adds a sentence to the rollback section of the production readiness answers. A rise in outcome="error", or a sustained rate of trigger="periodic",outcome="changed", means reconciliation is failing or Node updates are being missed. The README already notes the metric is only exposed by the CCM.
The Beta criteria were reworded to match. They now say the labels allow tracking corrections, rather than promising a separate gated metric.
The README also records a failure mode: previously a reconcile ran every 10s from a cached node list. Now the controller only reacts to events, so if the API server is unavailable, nothing happens. Plan your alerts for that, because a quiet counter can mean healthy or stuck.
What to watch ¶
- If you build dashboards for the CCM, key them on the
triggerandoutcomelabels, not on a corrections metric name. The old name will not exist. - Alert on a quiet
route_controller_route_sync_totalfor clusters with node churn, since the 10s heartbeat of reconciles goes away once the gate is on. - Follow the beta milestone in kep.yaml. Moving from v1.37 depends on the implementation PRs and the e2e items still unchecked in the KEP.
This is a documentation alignment, not a behavior change in a release. Its value is that the plan an operator reads now matches the code.