Rolling Updates

The operator allows changing the PostgreSQL version used in a cluster while applications are running against it.

Rolling upgrades are started when:

• the user changes the imageName attribute of the cluster specification;

• a change in the PostgreSQL configuration requires a restart to be applied;

• a change on the Cluster .spec.resources values

• a change in size of the persistent volume claim on AKS

• after the operator is updated, to ensure the Pods run the latest instance manager (unless in-place updates are enabled).

The operator starts upgrading all the replicas, one Pod at a time, and begins from the one with the highest serial.

The primary is the last node to be upgraded.

Rolling updates are configurable and can be either entirely automated (unsupervised) or requiring human intervention (supervised).

The upgrade keeps the CloudNativePG identity, without re-cloning the data. Pods will be deleted and created again with the same PVCs and a new image, if required.

During the rolling update procedure, each service endpoints move to reflect the cluster's status, so that applications can ignore the node that is being updated.

Automated updates (unsupervised)

When primaryUpdateStrategy is set to unsupervised, the rolling update process is managed by Kubernetes and is entirely automated. Once the replicas have been upgraded, the selected primaryUpdateMethod operation will initiate on the primary. This is the default behavior.

The primaryUpdateMethod option accepts one of the following values:

• switchover: a switchover operation is automatically performed, setting the most aligned replica as the new target primary, and shutting down the former primary pod (default).

• restart: if possible, perform an automated restart of the pod where the primary instance is running. Otherwise, the restart request is ignored and a switchover issued.

There's no one-size-fits-all configuration for the update method, as that depends on several factors like the actual workload of your database, the requirements in terms of RPO and RTO, whether your PostgreSQL architecture is shared or shared nothing, and so on.

Indeed, being PostgreSQL a primary/standby architecture database management system, the update process inevitably generates a downtime for your applications. One important aspect to consider for your context is the time it takes for your pod to download the new PostgreSQL container image, as that depends on your Kubernetes cluster settings and specifications. The switchover method makes sure that the promoted instance already runs the target image version of the container. The restart method instead might require to download the image from the origin registry after the primary pod has been shut down. It is up to you to determine whether, for your database, it is best to use restart or switchover as part of the rolling update procedure.

Manual updates (supervised)

When primaryUpdateStrategy is set to supervised, the rolling update process is suspended immediately after all replicas have been upgraded.

This phase can only be completed with either a manual switchover or an in-place restart. Keep in mind that image upgrades can not be applied with an in-place restart, so a switchover is required in such cases.

You can trigger a switchover with:

kubectl cnpg promote [cluster] [new_primary]

You can trigger a restart with:

kubectl cnpg restart [cluster] [current_primary]

You can find more information in the cnpg plugin page.