Postgres instance manager#

EDB CloudNativePG Cluster does not rely on an external tool for failover management. It simply relies on the Kubernetes API server and a native key component called: the Postgres instance manager.

The instance manager takes care of the entire lifecycle of the PostgreSQL server process (also known as postmaster ).

When you create a new cluster, the operator makes a Pod per instance. The field .spec.instances specifies how many instances to create.

Each Pod will start the instance manager as the parent process (PID 1) for the main container, which in turn runs the PostgreSQL instance. During the lifetime of the Pod, the instance manager acts as a backend to handle the startup, liveness and readiness probes .

Startup, liveness and readiness probes#

The startup and liveness probes rely on pg_isready , while the readiness probe checks if the database is up and able to accept connections.

Startup Probe#

The .spec.startDelay parameter specifies the delay (in seconds) before the liveness probe activates after a PostgreSQL Pod starts. By default, this is set to 3600 seconds. You should adjust this value based on the time PostgreSQL requires to fully initialize in your environment.

Warning

Setting .spec.startDelay too low can cause the liveness probe to activate prematurely, potentially resulting in unnecessary Pod restarts if PostgreSQL hasn’t fully initialized.

EDB CloudNativePG Cluster configures the startup probe with the following default parameters:

failureThreshold: FAILURE_THRESHOLD
periodSeconds: 10
successThreshold: 1
timeoutSeconds: 5

Here, FAILURE_THRESHOLD is calculated as startDelay divided by periodSeconds .

If the default behavior based on startDelay is not suitable for your use case, you can take full control of the startup probe by specifying custom parameters in the .spec.probes.startup stanza. Note that defining this stanza will override the default behavior, including the use of startDelay .

Warning

Ensure that any custom probe settings are aligned with your cluster’s operational requirements to prevent unintended disruptions.

Note

For detailed information about probe configuration, refer to the probe API .

For example, the following configuration bypasses startDelay entirely:

#  ... snip

spec:
  probes:
    startup:
      periodSeconds: 3
      timeoutSeconds: 3
      failureThreshold: 10

Liveness Probe#

The liveness probe begins after the startup probe succeeds and is responsible for detecting if the PostgreSQL instance has entered a broken state that requires a restart of the pod.

The amount of time before a Pod is classified as not alive is configurable via the .spec.livenessProbeTimeout parameter.

EDB CloudNativePG Cluster configures the liveness probe with the following default parameters:

failureThreshold: FAILURE_THRESHOLD
periodSeconds: 10
successThreshold: 1
timeoutSeconds: 5

Here, FAILURE_THRESHOLD is calculated as livenessProbeTimeout divided by periodSeconds .

By default, .spec.livenessProbeTimeout is set to 30 seconds. This means the liveness probe will report a failure if it detects three consecutive probe failures, with a 10-second interval between each check.

If the default behavior using livenessProbeTimeout does not meet your needs, you can fully customize the liveness probe by defining parameters in the .spec.probes.liveness stanza. Keep in mind that specifying this stanza will override the default behavior, including the use of livenessProbeTimeout .

Warning

Ensure that any custom probe settings are aligned with your cluster’s operational requirements to prevent unintended disruptions.

Note

For more details on probe configuration, refer to the probe API .

For example, the following configuration overrides the default behavior and bypasses livenessProbeTimeout :

#  ... snip

spec:
  probes:
    liveness:
      periodSeconds: 3
      timeoutSeconds: 3
      failureThreshold: 10

Readiness Probe#

The readiness probe determines when a pod running a PostgreSQL instance is prepared to accept traffic and serve requests.

EDB CloudNativePG Cluster uses the following default configuration for the readiness probe:

failureThreshold: 3
periodSeconds: 10
successThreshold: 1
timeoutSeconds: 5

If the default settings do not suit your requirements, you can fully customize the readiness probe by specifying parameters in the .spec.probes.readiness stanza. For example:

#  ... snip

spec:
  probes:
    readiness:
      periodSeconds: 3
      timeoutSeconds: 3
      failureThreshold: 10

Warning

Ensure that any custom probe settings are aligned with your cluster’s operational requirements to prevent unintended disruptions.

Note

For more information on configuring probes, see the probe API .

Shutdown control#

When a Pod running Postgres is deleted, either manually or by Kubernetes following a node drain operation, the kubelet will send a termination signal to the instance manager, and the instance manager will take care of shutting down PostgreSQL in an appropriate way. The .spec.smartShutdownTimeout and .spec.stopDelay options, expressed in seconds, control the amount of time given to PostgreSQL to shut down. The values default to 180 and 1800 seconds, respectively.

The shutdown procedure is composed of two steps:

  1. The instance manager requests a smart shut down, disallowing any new connection to PostgreSQL. This step will last for up to .spec.smartShutdownTimeout seconds.

  2. If PostgreSQL is still up, the instance manager requests a fast shut down, terminating any existing connection and exiting promptly. If the instance is archiving and/or streaming WAL files, the process will wait for up to the remaining time set in .spec.stopDelay to complete the operation and then forcibly shut down. Such a timeout needs to be at least 15 seconds.

Important

In order to avoid any data loss in the Postgres cluster, which impacts the database Purpose of Release Candidates , don't delete the Pod where the primary instance is running. In this case, perform a switchover to another instance first.

Shutdown of the primary during a switchover#

During a switchover, the shutdown procedure is slightly different from the general case. Indeed, the operator requires the former primary to issue a fast shut down before the selected new primary can be promoted, in order to ensure that all the data are available on the new primary.

For this reason, the .spec.switchoverDelay , expressed in seconds, controls the time given to the former primary to shut down gracefully and archive all the WAL files. By default it is set to 3600 (1 hour).

Warning

The .spec.switchoverDelay option affects the Purpose of Release Candidates

and RTO and RPO impact of your PostgreSQL database. Setting it to a low value, might favor RTO over RPO but lead to data loss at cluster level and/or backup level. On the contrary, setting it to a high value, might remove the risk of data loss while leaving the cluster without an active primary for a longer time during the switchover.

Failover#

In case of primary pod failure, the cluster will go into failover mode. Please refer to the Automated failover for self-healing for details.

Disk Full Failure#

Storage exhaustion is a well known issue for PostgreSQL clusters. The PostgreSQL documentation

highlights the possible failure scenarios and the importance of monitoring disk usage to prevent it from becoming full.

The same applies to EDB CloudNativePG Cluster and Kubernetes as well: the Monitoring and metrics

provides details on checking the disk space used by WAL segments and standard metrics on disk usage exported to Prometheus.

Important

In a production system, it is critical to monitor the database continuously. Exhausted disk storage can lead to a database server shutdown.

Note

The detection of exhausted storage relies on a storage class that accurately reports disk size and usage. This may not be the case in simulated Kubernetes environments like Kind or with test storage class implementations such as csi-driver-host-path .

If the disk containing the WALs becomes full and no more WAL segments can be stored, PostgreSQL will stop working. EDB CloudNativePG Cluster correctly detects this issue by verifying that there is enough space to store the next WAL segment, and avoids triggering a failover, which could complicate recovery.

That allows a human administrator to address the root cause.

In such a case, if supported by the storage class, the quickest course of action is currently to:

  1. Expand the storage size of the full PVC

  2. Increase the size in the Cluster resource to the same value

Once the issue is resolved and there is sufficient free space for WAL segments, the Pod will restart and the cluster will become healthy.

See also the Volume expansion of the documentation.