Installation and upgrades

Installation on Kubernetes

Directly using the operator manifest

The operator can be installed like any other resource in Kubernetes, through a YAML manifest applied via kubectl.

You can install the latest operator manifest for this minor release as follows:

kubectl apply -f \
  https://raw.githubusercontent.com/cloudnative-pg/cloudnative-pg/release-1.20/releases/cnpg-1.20.6.yaml

You can verify that with:

kubectl get deploy -n cnpg-system cnpg-controller-manager

Using the cnpg plugin for kubectl

You can use the cnpg plugin to override the default configuration options that are in the static manifests.

For example, to generate the default latest manifest but change the watch namespaces to only be a specific namespace, you could run:

kubectl cnpg install generate \
  --watch-namespace "specific-namespace" \
  > cnpg_for_specific_namespace.yaml

Please refer to "cnpg plugin" documentation for a more comprehensive example.

Testing the latest development snapshot

If you want to test or evaluate the latest development snapshot of CloudNativePG before the next official patch release, you can download the manifests from the cloudnative-pg/artifacts which provides easy access to the current trunk (main) as well as to each supported release.

For example, you can install the latest snapshot of the operator with:

curl -sSfL \
  https://raw.githubusercontent.com/cloudnative-pg/artifacts/main/manifests/operator-manifest.yaml | \
  kubectl apply -f -

If you are instead looking for the latest snapshot of the operator for this specific minor release, you can just run:

curl -sSfL \
  https://raw.githubusercontent.com/cloudnative-pg/artifacts/release-1.20/manifests/operator-manifest.yaml | \
  kubectl apply -f -

Using the Helm Chart

The operator can be installed using the provided Helm chart.

Details about the deployment

In Kubernetes, the operator is by default installed in the cnpg-system namespace as a Kubernetes Deployment. The name of this deployment depends on the installation method. When installed through the manifest or the cnpg plugin, it is called cnpg-controller-manager by default. When installed via Helm, the default name is cnpg-cloudnative-pg.

You can get more information using the describe command in kubectl:

$ kubectl get deployments -n cnpg-system
NAME                READY   UP-TO-DATE   AVAILABLE   AGE
<deployment-name>   1/1     1            1           18m

kubectl describe deploy \
  -n cnpg-system \
  <deployment-name>

As with any Deployment, it sits on top of a ReplicaSet and supports rolling upgrades. The default configuration of the CloudNativePG operator comes with a Deployment of a single replica, which is suitable for most installations. In case the node where the pod is running is not reachable anymore, the pod will be rescheduled on another node.

If you require high availability at the operator level, it is possible to specify multiple replicas in the Deployment configuration - given that the operator supports leader election. Also, you can take advantage of taints and tolerations to make sure that the operator does not run on the same nodes where the actual PostgreSQL clusters are running (this might even include the control plane for self-managed Kubernetes installations).

Upgrades

Upgrading CloudNativePG operator is a two-step process:

1. upgrade the controller and the related Kubernetes resources
2. upgrade the instance manager running in every PostgreSQL pod

Unless differently stated in the release notes, the first step is normally done by applying the manifest of the newer version for plain Kubernetes installations, or using the native package manager of the used distribution (please follow the instructions in the above sections).

The second step is automatically executed after having updated the controller, by default triggering a rolling update of every deployed PostgreSQL instance to use the new instance manager. The rolling update procedure culminates with a switchover, which is controlled by the primaryUpdateStrategy option, by default set to unsupervised. When set to supervised, users need to complete the rolling update by manually promoting a new instance through the cnpg plugin for kubectl.

Since version 1.10.0, the rolling update behavior can be replaced with in-place updates of the instance manager. The latter don't require a restart of the PostgreSQL instance and, as a result, a switchover in the cluster. This behavior, which is disabled by default, is described below.

In-place updates of the instance manager

By default, CloudNativePG issues a rolling update of the cluster every time the operator is updated. The new instance manager shipped with the operator is added to each PostgreSQL pod via an init container.

However, this behavior can be changed via configuration to enable in-place updates of the instance manager, which is the PID 1 process that keeps the container alive.

Internally, any instance manager from version 1.10 of CloudNativePG supports injection of a new executable that will replace the existing one, once the integrity verification phase is completed, as well as graceful termination of all the internal processes. When the new instance manager restarts using the new binary, it adopts the already running postmaster.

As a result, the PostgreSQL process is unaffected by the update, refraining from the need to perform a switchover. The other side of the coin, is that the Pod is changed after the start, breaking the pure concept of immutability.

You can enable this feature by setting the ENABLE_INSTANCE_MANAGER_INPLACE_UPDATES environment variable to 'true' in the operator configuration.

The in-place upgrade process will not change the init container image inside the Pods. Therefore, the Pod definition will not reflect the current version of the operator.

Compatibility among versions

CloudNativePG follows semantic versioning. Every release of the operator within the same API version is compatible with the previous one. The current API version is v1, corresponding to versions 1.x.y of the operator.

In addition to new features, new versions of the operator contain bug fixes and stability enhancements. Because of this, we strongly encourage users to upgrade to the latest version of the operator, as each version is released in order to maintain the most secure and stable Postgres environment.

CloudNativePG currently releases new versions of the operator at least monthly. If you are unable to apply updates as each version becomes available, we recommend upgrading through each version in sequential order to come current periodically and not skipping versions.

The release notes page contains a detailed list of the changes introduced in every released version of CloudNativePG, and it must be read before upgrading to a newer version of the software.

Most versions are directly upgradable and in that case, applying the newer manifest for plain Kubernetes installations or using the native package manager of the chosen distribution is enough.

When versions are not directly upgradable, the old version needs to be removed before installing the new one. This won't affect user data but only the operator itself.

Upgrading to 1.22.0, 1.21.2 or 1.20.5

CloudNativePG continues to adhere to the security-by-default approach. As of version 1.22, the usage of the ALTER SYSTEM command is now disabled by default.

The reason behind this choice is to ensure that, by default, changes to the PostgreSQL configuration in a database cluster controlled by CloudNativePG are allowed only through the Kubernetes API.

At the same time, we are providing an option to enable ALTER SYSTEM if you need to use it, even temporarily, from versions 1.22.0, 1.21.2, and 1.20.5, by setting .spec.postgresql.enableAlterSystem to true, as in the following excerpt:

...
  postgresql:
    enableAlterSystem: true
...

Clusters in 1.22 will have enableAlterSystem set to false by default. If you want to retain the existing behavior, in 1.22, you need to explicitly set enableAlterSystem to true as shown above.

In versions 1.21.2 and 1.20.5, and later patch releases in the 1.20 and 1.21 branches, enableAlterSystem will be set to true by default, keeping with the existing behavior. If you don't need to use ALTER SYSTEM, we recommend that you set enableAlterSystem explicitly to false.

Upgrading to 1.20.3

With the goal to keep improving out-of-the-box the convention over configuration behavior of the operator, CloudNativePG changes the default value of several knobs in the following areas:

• startup and shutdown control of the PostgreSQL instance
• self-healing
• labels

Delay for PostgreSQL shutdown

Up to 1.20.2, the stopDelay parameter was set to 30 seconds. Despite the recommendations to change and tune this value, almost all the cases we have examined during support incidents or community issues show that this value is left unchanged.

The new default value is 1800 seconds, the equivalent of 30 minutes.

The new smartShutdownTimeout parameter has been introduced to define the maximum time window within the stopDelay value reserved to complete the smart shutdown procedure in PostgreSQL. During this time, the Postgres server rejects any new connections while waiting for all regular sessions to terminate.

Once elapsed, the remaining time up to stopDelay will be reserved for PostgreSQL to complete its duties regarding WAL commitments with both the archive and the streaming replicas to ensure the cluster doesn't lose any data.

If you want to retain the old behavior, you need to set explicitly:

spec:
   ...
   stopDelay: 30

And, after the upgrade has completed, specify smartShutdownTimeout:

spec:
   ...
   stopDelay: 30
   smartShutdownTimeout: 15

Delay for PostgreSQL startup

Up to 1.20.2, the startDelay parameter was set to 30 seconds, and CloudNativePG used this parameter as initialDelaySeconds for the Kubernetes liveness probe. Given that all the supported Kubernetes releases provide startup probes, startDelay is now automatically divided into periods of 10 seconds of duration each.

Despite the recommendations to change and tune this value, almost all the cases we have examined during support incidents or community issues show that this value is left unchanged. Given that this parameter influences the startup of a PostgreSQL instance, a low value of startDelay would cause Postgres never to reach a consistent recovery state and be restarted indefinitely.

For this reason, startDelay has been raised by default to 3600 seconds, the equivalent of 1 hour.

If you want to retain the existing behavior using the new implementation, you can do that by explicitly setting:

spec:
   ...
   startDelay: 30

Delay for PostgreSQL switchover

Up to 1.20.2, the switchoverDelay parameter was set by default to 40000000 seconds (over 15 months) to simulate a very long interval.

The default value has been lowered to 3600 seconds, the equivalent of 1 hour.

If you want to retain the old behavior, you need to set explicitly:

spec:
   ...
   switchoverDelay: 40000000

Labels

In version 1.18, we deprecated the postgresql label in pods to identify the name of the cluster, and replaced it with the more canonical cnpg.io/cluster label. The postgresql label is no longer maintained.

Similarly, from this version, the role label is deprecated. The new label cnpg.io/instanceRole is now used, and will entirely replace the role label in a future release.

Shortcut for keeping the existing behavior

If you want to explicitly keep the behavior of CloudNativePG up to version 1.20.2 (we advise not to), you need to set these values in all your Cluster definitions before upgrading to a higher version:

spec:
   ...
   startDelay: 30
   stopDelay: 30
   switchoverDelay: 40000000

Once the upgrade is completed, also add:

spec:
   ...
   smartShutdownTimeout: 15

Upgrading to 1.20 from a previous minor version

CloudNativePG 1.20 introduces some changes from previous versions of the operator in the default behavior of a few features, with the goal to improve resilience and usability of a Postgres cluster out of the box, through convention over configuration.

Backup from a standby

Backup from a standby was introduced in CloudNativePG 1.19, but disabled by default - meaning that the base backup is taken from the primary unless the target is explicitly set to prefer standby.

From version 1.20, if one or more replicas are available, the operator will prefer the most aligned standby to take a full base backup.

If you are upgrading your CloudNativePG deployment to 1.20 and are concerned that this feature might impact your production environment for the new Cluster resources that you create, you can explicitly set the target to the primary by adding the following line to all your Cluster resources:

spec:
   ...
   backup:
     target: "primary"

Restart of a primary after a rolling update

Automated rolling updates have been always available in CloudNativePG, and by default they update the primary after having performed a switchover to the most aligned replica.

From version 1.20, we are changing the default update method of the primary from switchover to restart as, in most cases, this is the fastest and safest way.

If you are upgrading your CloudNativePG deployment to 1.20 and are concerned that this feature might impact your production environment for the new Cluster resources that you create, you can explicitly set the update method of the primary to switchover by adding the following line to all your Cluster resources:

spec:
   ...
   primaryUpdateMethod: switchover

Replication slots for High Availability

Replication slots for High Availability were introduced in CloudNativePG in version 1.18, but disabled by default.

Version 1.20 prepares the ground for enabling this feature by default in any future release, as replication slots enhance the resilience and robustness of a High Availability cluster.

For future compatibility, if you already know that your environments won't ever need replication slots, our recommendation is that you explicitly disable their management by adding from now the following lines to your Cluster resources:

spec:
   ...
   replicationSlots:
     highAvailability:
       enabled: false