Workload Configuration

This document describes annotations and labels that you can manually configure on your workloads (Deployment / StatefulSet / DaemonSet) to customize the behavior of the Workload Autoscaler.

Disable Proactive Update

You can disable proactive updates (eviction / in-place resize) for a specific workload by adding the following annotation:

Example:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: my-app
  annotations:
    evpa.cloudpilot.ai/disable-proactive-update: "true"

Note: When this annotation is set to "true", all optimization actions are disabled for this workload, including applying recommendations to newly created Pods. Recommendations are still calculated in the background and will take effect once you remove the annotation or set it to "false".

Override Runtime Language

CloudPilot AI automatically detects each container’s runtime language via eBPF. If the auto-detected runtime language is not correct for your workload, you can override it using a label:

Example: override the app container’s runtime language to java:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: my-app
  labels:
    evpa.cloudpilot.ai/container-runtime-language.app: "java"

This is useful when:

• The eBPF agent has not yet detected the language (e.g., workload just started).
• The auto-detected language is incorrect for a specific container.

Override Primary Container

CloudPilot AI automatically selects a primary container for each workload based on the container with the highest total resource requests (CPU + Memory). If no container has resource requests set, the first container in the Pod spec is used.

You can override this selection with a label:

Example:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: my-app
  labels:
    evpa.cloudpilot.ai/primary-container: "app"

The primary container is used to determine:

• The workload’s primary runtime language label (evpa.cloudpilot.ai/primary-runtime-language), which can be used in AutoscalingPolicy LabelSelector to target workloads by language.
• Whether to apply Java-specific optimization (JVM Heap management, GC-aware recommendations, etc.) when the primary container’s runtime language is java.

System-Managed Labels (Read-Only)

The following labels are automatically managed by the Workload Autoscaler. You should not modify them manually — they are listed here for reference only.

Note: The evpa.cloudpilot.ai/primary-runtime-language label is useful for targeting workloads in AutoscalingPolicy via LabelSelector. For example, you can create a policy that applies only to Java workloads by selecting evpa.cloudpilot.ai/primary-runtime-language: java. See AutoscalingPolicy LabelSelector for details.

Runtime Telemetry and Optimization

Disabling runtime optimization does not disable reporting of available JVM, usage or pressure data. Independent successful fields remain reportable when another query fails. Missing observations stay absent and measured zero stays present. Display recommendations can fall back to the original Request when a target is missing; that display fallback is not an executable recommendation.

For selected workloads, each resource axis requires observation coverage for the selected members. For native workloads, reported usage totals sum the Pods that were actually observed; missing Pod samples are not extrapolated across every replica. Upgrade the CloudPilot backend before the Agent so that charts can consume these observed totals. Older backends continue to use the legacy per-Pod average and replica calculation.