Minor Rollback

Example: PgBouncer 1.19 -> PgBouncer 1.18
(including simple charm revision bump: from revision 43 to revision 42)

WARNING: do NOT trigger rollback during the running upgrade action! It may cause unpredictable PostgreSQL Cluster and/or PgBouncer state!

Minor rollback steps

The rollback is NOT necessary if sacrificial unit is created before the upgrade, just remove it using scale-down the application. Otherwise perform the rollback:

1. Rollback. Perform the charm rollback using juju refresh. The unit with the maximal ordinal will be rolled-back first and rollback continues for entire application.
2. Check. Make sure the charm and cluster are in healthy state again.

Manual Rollback

After a juju refresh, case there any version incompatibilities in charm revisions or it dependencies, or any other unexpected failure in the upgrade process, the upgrade process will be halted an enter a failure state.

Although the underlying PostgreSQL Cluster and PgBouncer continue to work, it’s important to rollback the charm to previous revision so an update can be later attempted after a further inspection of the failure.

To execute a rollback we take the same procedure as the upgrade, the difference being the charm revision to upgrade to. In case of this tutorial example, one would refresh the charm back to revision 88, the steps being:

Step 1: Rollback

When using charm from charmhub:

juju refresh pgbouncer-k8s --revision=88 --trust

Case deploying from local charm file, one need to have the previous revision charm file and the postgresql-image resource, then run:

juju refresh pgbouncer-k8s --trust --path=./pgbouncer-k8s_ubuntu-22.04-amd64.charm \
       --resource pgbouncer-image=ghcr.io/canonical/charmed-postgresql@sha256:753477ce39712221f008955b746fcf01a215785a215fe3de56f525380d14ad97

Where pgbouncer-k8s_ubuntu-22.04-amd64.charm is the previous revision charm file. The reference for the resource for a given revision can be found at the metadata.yaml file in the charm repository.

The biggest ordinal unit will be rolled out and should rejoin the cluster after settling down. After the refresh command, the juju controller revision for the application will be back in sync with the running PgBouncer K8s revision.

Step 2: Check

The future improvement is planned to check the state on pod/cluster on a low level. At the moment check juju status to make sure the cluster state is OK.