The GitHub Actions job "Build Docker Image" on 
flink-kubernetes-operator.git/FLINK-40407-scaling-reason-tags has failed.
Run started by GitHub user kaustubhbutte17 (triggered by kaustubhbutte17).

Head commit for run:
f1aa68741d1f64e16725eb93cdfcdf6997b45145 / Kaustubh Butte 
<[email protected]>
[FLINK-40407] Add reason tags to the autoscaler.scalings counter

The autoscaler.scalings counter reports that a cycle changed the parallelism.
It does not report what drove the change. An operator cannot tell a backlog
driven scale up from a sustained load one, and cannot alert on only the kind
that matters.

This commit adds a ScaleReason enum with four values: backlog, high_load,
input_spike and low_util. AutoscalerFlinkMetrics reports a second counter that
carries the reasons as a metric variable. The untagged scalings counter keeps
the total, so existing dashboards and alerts continue to work.

JobVertexScaler names the reasons next to the decision, and reads the same
values that the target capacity formula uses. A reason that is derived
somewhere else drifts away from the algorithm as the algorithm changes.

A vertex carries a set and not one value, because the target capacity adds the
lag catch up rate to the rate at the target utilization. A backlog and a high
load can therefore apply to the same decision. The reasons of every scaled
vertex join into one sorted tag, for example backlog|high_load. Four reasons
give at most 15 values.

ScalingSummary carries the reasons but does not serialize them. The scaling
history in the state store keeps the shape it had before, so an operator can
roll back without a state migration.

Report URL: 
https://github.com/apache/flink-kubernetes-operator/actions/runs/37021060585

With regards,
GitHub Actions via GitBox

Reply via email to