Configure CPU and Memory for Confluent Platform in Confluent for Kubernetes

Setting proper requests and limits is important for the performance of the Confluent Platform clusters and their applications.

Before configuring CPU and memory resource requirements for Confluent Platform, review Cluster sizing for resource allocation planning.

Specify CPU and memory requests

Requests and limits are dependent on your workload. The best practice should follow the minimum requirements for Confluent Platform, and then benchmark and tune if needed to suit your environment.

For more information about CPU and memory resources in Kubernetes, see Resource Management for Pods and Containers.

Confluent for Kubernetes (CFK) allows you to define custom pod resource requirements for Confluent Platform components it deploys. You specify these requirements using the requests and limits properties for components in their custom resources (CR).

spec:
  podTemplate:
    resources:  --- [1]
      limits:   --- [2]
        cpu:    --- [3]
        memory: --- [4]
      requests: --- [5]
        cpu:    --- [6]
        memory: --- [7]

  • [1] resources describe the compute resource requirements for this component CR.

  • [2] limits describe the maximum amount of compute resources allowed. Your Confluent Platform component will throttle if it tries to use more resources than the values set here.

  • [3] [6] Limits and requests for CPU resources are measured in CPU units. 1 CPU unit is equivalent to 1 physical CPU core or 1 virtual core.

    Fractional CPU requests are allowed. For example, resources.requests.cpu: 0.5 requests half as much CPU time.

    For CPU resource units, the quantity expression 0.1 is equivalent to the expression 100m, which means “one hundred milli CPU”.

  • [4] [7] Limits and requests for memory are measured in bytes.

  • [5] requests describe the minimum amount of compute resources required. If the requests section is omitted, it defaults to limits if that is explicitly specified in the same CR, and otherwise to the values defined for the Kubernetes cluster.

The following example CR specifies requests for 0.25 CPU and 64 MiB of memory. The limits are set to 0.5 CPU and 128 MiB of memory.

spec:
  podTemplate:
    resources:
      requests:
        memory: "64Mi"
        cpu: "250m"
      limits:
        memory: "128Mi"
        cpu: "500m"

Resource definitions for CFK Init Container

CFK does not allow you to configure the resource requests or limits for the Init Container. For each Confluent Platform deployment, CFK sets the following for the Init Container:

resources:
  limits:
    cpu: 500m
    memory: 1Gi
  requests:
    cpu: 100m
    memory: 512Mi

Define Java heap size

In addition to memory sizing, you can configure the Java (JVM) heap size for Confluent components using the configuration override feature (spec.configOverrides.jvm).

For guidance on setting JVM heap size, see the Tuning defaults for containers article.

Auto scale JVM heap size

To auto scale the JVM heap size, make sure UseContainerSupport is enabled, which is the default setting:

kind: Kafka
spec:
  configOverrides:
    jvm:
      - "-XX:+UseContainerSupport"

Set JVM heap size

In the component CR, set MaxRAMPercentage to match the desired JVM heap size in relation to the total memory limit (spec.resource.limits.memory) you set in Specify CPU and memory requests above.

kind: Kafka
spec:
  configOverrides:
    jvm:
      - "-XX:MaxRAMPercentage=<percentage value>"

To use MaxRAMPercentage, you must first remove the existing JVM -Xms and -Xmx settings, using the current memory limits or requests for the pod. When only the pod memory limits value is set, use the memory limits value to remove the setting. When both the memory limits and the requests are set, use the requests value to remove the settings. See below for examples.

For the steps to remove configOverrides settings, see Remove default configuration.

Note

To remove -Xms and -Xmx, specify the value in megabytes (M) or gigabytes (G). Mebibytes (Mi) or gibibytes (Gi) are not supported.

The following examples show the JVM MaxRAMPercentage set to 50% of the memory and auto-scaling enabled.

Example 1: In this example, only limits is set (4Gi) for the pod. To remove the -Xmx and -Xms, specify the pods memory limit, such as ---Xmx4G and ---Xms4G.

apiVersion: platform.confluent.io/v1beta1
kind: Kafka
metadata:
  name: kafka
  namespace: confluent
spec:
  podTemplate:
    resources:
       limits:
         memory: "4Gi"

  configOverrides:
    jvm:
      - "---Xmx4G" # This removes -Xmx parameter
      - "---Xms4G" # This removes -Xms parameter
      - "-XX:+UseContainerSupport"
      - "-XX:MaxRAMPercentage=50.0"

Example 2: In this example both limits and requests are set for the pod. To remove the -Xmx and -Xms, use the pods requests memory, such as ---Xmx2G and ---Xms2G.

apiVersion: platform.confluent.io/v1beta1
kind: Kafka
metadata:
  name: kafka
  namespace: confluent
spec:
  podTemplate:
    resources:
       limits:
         memory: "4Gi"
       requests:
         memory: "2Gi"

  configOverrides:
    jvm:
      - "---Xmx2G" # This removes -Xmx parameter
      - "---Xms2G" # This removes -Xms parameter
      - "-XX:+UseContainerSupport"
      - "-XX:MaxRAMPercentage=50.0"

For more details about the configuration overrides feature, see Configuration overrides.