Resource management

In a typical Kubernetes cluster, pods run with unlimited resources. By default, they might be allowed to use as much CPU and RAM as needed.

Cloud Native PostgreSQL allows administrators to control and manage resource usage by the pods of the cluster, through the resources section of the manifest, with two knobs:

  • requests: initial requirement
  • limits: maximum usage, in case of dynamic increase of resource needs

For example, you can request an initial amount of RAM of 32MiB (scalable to 128MiB) and 50m of CPU (scalable to 100m) as follows:

      memory: "32Mi"
      cpu: "50m"
      memory: "128Mi"
      cpu: "100m"

Memory requests and limits are associated with containers, but it is useful to think of a pod as having a memory request and limit. The pod's memory request is the sum of the memory requests for all the containers in the pod.

Pod scheduling is based on requests and not on limits. A pod is scheduled to run on a Node only if the Node has enough available memory to satisfy the pod's memory request.

For each resource, we divide containers into 3 Quality of Service (QoS) classes, in decreasing order of priority:

  • Guaranteed
  • Burstable
  • Best-Effort

For more details, please refer to the "Configure Quality of Service for Pods" section in the Kubernetes documentation.

For a PostgreSQL workload it is recommended to set a "Guaranteed" QoS.

To avoid resources related issues in Kubernetes, we can refer to the best practices for "out of resource" handling while creating a cluster:

  • Specify your required values for memory and CPU in the resources section of the manifest file. This way, you can avoid the OOM Killed (where "OOM" stands for Out Of Memory) and CPU throttle or any other resources related issues on running instances.
  • For your cluster's pods to get assigned to the "Guaranteed" QoS class, you must set limits and requests for both memory and CPU to the same value.
  • Specify your required PostgreSQL memory parameters consistently with the pod resources (as you would do in a VM or physical machine scenario - see below).
  • Set up database server pods on a dedicated node using nodeSelector. See the "nodeSelector field of the affinityconfiguration resource on the API reference page".

You can refer to the following example manifest:

kind: Cluster
  name: postgresql-resources

  instances: 3

      shared_buffers: "256MB"

      memory: "1024Mi"
      cpu: 1
      memory: "1024Mi"
      cpu: 1

    size: 1Gi

In the above example, we have specified shared_buffers parameter with a value of 256MB - i.e., how much memory is dedicated to the PostgreSQL server for caching data (the default value for this parameter is 128MB in case it's not defined).

A reasonable starting value for shared_buffers is 25% of the memory in your system. For example: if your shared_buffers is 256 MB, then the recommended value for your container memory size is 1 GB, which means that within a pod all the containers will have a total of 1 GB memory that Kubernetes will always preserve, enabling our containers to work as expected. For more details, please refer to the "Resource Consumption" section in the PostgreSQL documentation.

Managing Compute Resources for Containers

For more details on resource management, please refer to the "Managing Compute Resources for Containers" page from the Kubernetes documentation.