The CRDs come first because the KubeBlocks operator and the chart create
objects of the kinds they define, and they can’t ship inside the chart: the
KubeBlocks CRDs are 11 MB, and Helm keeps each release in a Secret capped at
about 1 MB.
install.sh applies them for you; with Argo CD or Flux, the
falkordb-enterprise-crds bundle does. It holds the same CRD versions the
installer and the chart’s CRD hook apply, as plain manifests in two
directories, each with a kustomization.yaml:
The bundle and the chart share a version number. Pin both to the same one.
Before you start
install.sh sets up a few things that you now set up yourself. Keep the
namespaces in Git. Never commit the Secrets’ values: create the Secrets with
kubectl, or through a secrets operator such as External Secrets or Sealed
Secrets.
- Namespaces. The release namespace (
falkordb-systembelow) andkb-system, with the Pod Security labels in Pod security. - Registry credentials. A
kubernetes.io/dockerconfigjsonSecret forregistry.falkordb.cloudin the release namespace, named inglobal.imageCredentials.namewithcreate: false, and one your GitOps controller uses to pull the bundle and the chart. - The Admin Server’s JWT signing secret. A Secret in the release namespace,
named in
adminServer.secret.jwtExistingSecret.
Values for the chart
Whichever controller you use, give the chart these values on top of your own:install.sh gives it:
Volume snapshots
Volume-snapshot backups need theVolumeSnapshot CRDs, a snapshot controller,
and a VolumeSnapshotClass for your CSI driver. install.sh sets these up
from what it finds in the cluster; with Argo CD or Flux, choose yourself:
- Your cluster already runs a snapshot controller (most managed
Kubernetes services do). Set
volumeSnapshotClass.enabled: true, withvolumeSnapshotClass.driverset to your CSI driver, or leave itfalseand use aVolumeSnapshotClassyou already have. - It doesn’t. Set
snapshot-controller.enabled: trueand leavesnapshot-controller.installCRDsat its default,true, plus theVolumeSnapshotClassvalues above. The chart then installs the controller, its conversion webhook and the CRDs together, with the CRDs pointing at that webhook and its CA. Don’t apply the bundle’svolume-snapshotsdirectory as well. - You run your own snapshot controller without its CRDs. Apply the
bundle’s
volume-snapshotsdirectory in its own Application orKustomizationnext tokubeblocks. ItsVolumeGroupSnapshotContentCRD converts through a webhook, and names the upstream placeholder Service (default/snapshot-conversion-webhook-service) with no CA, so patch itsspec.conversion.webhook.clientConfigto your webhook from Git as well, for example with akustomization.yamlpatch, or conversions fail. - You don’t take volume-snapshot backups. Keep
volumeSnapshotClass.enabled: false.
VolumeSnapshotClass in the chart fails the sync, since
its kind doesn’t exist.
Argo CD
Argo CD 3.1 or later reads the bundle straight from the OCI registry. Register the registry once, as a repository Secret in Argo CD’s namespace. Like the other Secrets, create it outside Git:ignoreDifferences skip only empty values that KubeBlocks adds to or drops
from the chart’s ComponentDefinition, ActionSet and OpsDefinition
objects once they’re applied, such as resources: {}. They mean the same
either way, but without these rules the Application never shows as synced.
The chart is a Helm OCI repository: register registry.falkordb.cloud/falkordb
as a Helm repository with enableOCI: "true" and the same credentials.
The sync waves order the three Applications only when an app of apps creates
them, and only once Argo CD reports an Application’s health, which it doesn’t
by default; see Argo CD’s
sync waves
documentation. Otherwise, create them in order the first time, and wait for
each to sync before the next. Until the CRDs exist, the other two fail their
dry run with the server could not find the requested resource.
Argo CD renders the chart with helm template, so the chart’s hooks become
Argo CD hooks that run on every sync. That is why the values above keep the
CRD hook off. The chart’s other checks that look the cluster up, such as the
guards on the Admin Server’s ServiceAccount, can’t run under helm template
either.
Flux
Flux reads the bundle as anOCIRepository and applies it with a
Kustomization. The bundle’s layer is a plain tar archive, so select it by
its media type:
HelmRelease objects in a second Kustomization that
depends on the first, so they wait for the CRDs:
Upgrade
Read the changelog for the versions you skip first, as for any upgrade. Then upgrade in two commits, asinstall.sh does:
1
Upgrade the CRDs
Change the bundle’s version to the new chart version, and wait until the
CRDs are applied: the Argo CD Application is synced, or the Flux
Kustomization reports the new revision as its last applied one. The new
CRDs serve the KubeBlocks and chart you run now.2
Upgrade KubeBlocks and the chart
Change the chart to the new version, and KubeBlocks to the version
Compatibility lists for it.
dependsOn wait until the CRDs are healthy or ready, which the old CRDs
already are, so KubeBlocks or the chart can start upgrading before the new
CRDs are applied.
Air-gapped registries
Copy the bundle into your registry next to the chart and images, for example with ORAS:OCIRepository at your copy. See
Private registries for the chart and images.
The bundle is signed like the chart; verify it as
Supply chain describes for the chart.