Apply and manage Pod rightsizing for K8s Jobs with CLI

This topic describes how to apply Pod rightsizing to Kubernetes Jobs using the JobPilotConfig (JPC) system. You define JPC configurations in YAML and apply them to your cluster with kubectl.

To apply and manage Pod rightsizing on other workload types, see Apply and manage Pod rightsizing with CLI.

Kubernetes does not group Jobs under a single workload identity. Because each Job run creates its own Pod with an independent lifecycle, Kompass uses label selectors in the JobPilotConfig to collect related Jobs under a single Job group. A Job group represents all Jobs that share the same labels, allowing Kompass to analyze their behavior collectively and generate accurate rightsizing recommendations.

Workload selection components

Label-based selection (appSelector.labels)

The appSelector.labels field defines label-based rules for selecting Jobs to include in the Job group.

appSelector:
  labels:
    matchLabels:
      <label_key>: <label_value>

Behavior

• Includes all Jobs that match the specified labels.

• Treats all Jobs in the group as instances of the same workload for optimization purposes.

Label selector caution

Job grouping depends entirely on labels. If label selectors are too broad, they may group unrelated Jobs together. If they are too narrow, they may exclude Jobs from the group.

Priority configuration

You can assign a priority to a JPC to determine which configuration manages a Job when multiple JPCs match the same Job.

The JPC with the higher priority (lower numeric value) takes precedence and manages the matching Jobs.

Priority values start at 0 (highest priority).

Optimization strategy configuration

The optimization strategy defines how Kompass calculates CPU and memory recommendations for containers within the Job group.

Default strategy

Reduces CPU or memory to the lowest safe level for maximum savings. Appropriate for most Jobs.

• CPU: P95 on container usage over lookback period + configured buffer

• Memory: P99 on container usage over lookback period + configured buffer

Conservative strategy

Reduces CPU or memory moderately for moderate savings. Appropriate for sensitive or stateful Jobs.

• CPU: P99 on container usage over lookback period + configured buffer

• Memory: P100 on container usage over lookback period + configured buffer

For more information about optimization strategies, see Pod rightsizing workload optimization.

How Pod rightsizing applies to Jobs

Pod rightsizing for Jobs works differently than for long-running workloads because Jobs have a defined lifecycle with a start and end.

Kompass applies rightsizing changes only after a Job completes. Changes never interrupt running Jobs, ensuring zero disruption to your batch workloads. When a Job in the group runs again, it uses the new optimized resource requests.

Rollback protection

If a Job experiences CPU throttling or out-of-memory (OOM) events with the optimized resource requests, Kompass automatically rolls back the optimization for the next Job run.

The rollback sets the resource request to the maximum observed usage plus the configured buffer. This value is capped by the container's limit (if set). You can disable rollback for a Job group using the ignoreCPUThrottlingEvent and ignoreOOMEvent parameters.

Configuration template

Not all fields are configurable. Everything above the metadata section is predefined and should not be changed.

Values inside <> are user-specified placeholders.

...
metadata:
  name: <>
spec:
  appSelector:
    labels:
      matchLabels:
        <>
  cooldownPeriod: <>
  dataCoverage: <>
  ignoreCPUThrottlingEvent: <>
  ignoreOOMEvent: <>
  lookbackPeriod: <>
  managed: <>
  metrics:
    cpu:
      buffer: <>
      metricStrategy: <>
    memory:
      buffer: <>
      metricStrategy: <>
  minimumImpact: <>
  priority: <> 

Parameter reference

All parameters in the JobPilotConfig are mandatory.

Deploy the Pod rightsizing for Jobs configuration

Prerequisites

Pod rightsizing is enabled on the cluster. For more information, see Enable Pod rightsizing.

To deploy the JobPilotConfig:

1. Save the configuration to a YAML file and apply it to your cluster:

   kubectl apply -f jpc.yaml -n <namespace>

   Replace <namespace> with the namespace where your Jobs run.

2. (Optional) Verify that the JPC resource was created:

   kubectl get jpc

   Kompass activates Pod rightsizing for all matched Jobs within approximately one hour.

You can inspect or update the configuration using standard kubectl describe, kubectl apply, or kubectl delete commands.

Remove Pod rightsizing from a Job group

After Pod rightsizing is active on a Job group, you can remove the optimization. Removal takes effect within an hour.

To remove Pod rightsizing from a Job group:

Delete the JobPilotConfig resource:

kubectl delete jpc <jpc-name> -n <namespace>

Where:

• <jpc-name> is the name of your JobPilotConfig

• <namespace> is the namespace where it was deployed

Change the current optimization strategy

After Pod rightsizing is active on a Job group, you can change the current optimization strategy. Changes take effect within an hour.

To change an optimization strategy:

1. Edit the JobPilotConfig:

   kubectl edit jpc <jpc-name> -n <namespace>

2. Update the metrics.cpu.metricStrategy and/or metrics.memory.metricStrategy fields to your preferred strategy (default or conservative).

3. Save the changes.

   Kompass applies the new strategy within approximately one hour.

Troubleshooting

• No recommendations appear: The Job may not have enough runs to satisfy lookbackPeriod and dataCoverage.

• Job runs infrequently: Consider adjusting the lookbackPeriod.

• Job group not visible: Verify that the labels in matchLabels match the Job manifest.