Redpanda Operator Release Notes

This page lists the changes in each Redpanda Operator release in this release line, newest first. Each release shows what to know before you upgrade and the command to upgrade to that version.

Every operator release changes the sidecar image that runs in the broker Pods, so upgrading the operator restarts the brokers of Redpanda and StretchCluster resources. The Upgrade notes column says whether a release has anything else to know before you upgrade.

For a summary of the major features in this release line, see What’s New in the Redpanda Operator. To check which versions of Redpanda, Kubernetes, and Helm each release supports, see Kubernetes Compatibility.

Version Highlights Upgrade notes

v26.3.1
2026-11-20

RedpandaBrokerPool replaces NodePool, Redpanda Connect support is now GA, and Broker resources (beta) let a cluster move its brokers from StatefulSets to per-broker Broker resources with no restarts.

Yes

v26.3.1 (2026-11-20)

RedpandaBrokerPool replaces NodePool, Redpanda Connect support is now GA, and Broker resources (beta) let a cluster move its brokers from StatefulSets to per-broker Broker resources with no restarts.

Before you upgrade
  • The first upgrade to this release rolls brokers once if their Pod template changed since it was last applied.

  • The upgrade rolls brokers once to drop the helm.sh/chart label from their Pod templates.

  • The configurator and fs-validator volumes are renamed, so the upgrade rolls the brokers. Re-key any podTemplate override that uses the old <fullname>- prefixed volume name.

  • StretchCluster brokers restart once: their configurator init container gains a security context, and their startup and liveness probes set timeoutSeconds: 5.

  • Brokers with a trustStore on an internal Admin API, Kafka API, or Schema Registry listener restart once.

  • Clusters that use Broker resources roll every broker once to add the detect-disk-lost init container.

  • Rename the operator chart’s commonAnnotations value to connectAnnotations, and the --common-annotations flag to --connect-annotations. The old name fails helm upgrade.

  • NodePools that leave image unset roll to the Redpanda resource’s image. If that is older than the version they run, set image.tag on the Redpanda resource or the NodePool to the current version first.

  • The decommissioner started by --additional-controllers=decommission (or all) deletes claims only when --auto-delete-pvcs is passed.

  • The operator refuses to start with --additional-controllers=legacy-decommission. Use --additional-controllers=decommission instead.

  • If you override the multicluster ServiceMonitor to scheme: http, drop the override, or pass --metrics-secure=false to keep the plain-HTTP endpoint.

  • StretchCluster pools that set external.domain or external.addresses advertise a new external address after the next broker restart, so create DNS records for those names.

  • The operator ignores spec.clusterSpec.console and no longer updates a Console resource it created from it. Manage that Console resource directly.

  • The default Redpanda image tag (--redpanda-tag) is now v26.3.1. Pipeline resources that do not set an image move to Redpanda Connect 4.119.0, and Console resources move to Redpanda Console v3.15.0.

Install or upgrade to this version
helm repo add redpanda https://charts.redpanda.com
helm repo update
helm upgrade --install redpanda-controller redpanda/operator \
  --namespace <namespace> \
  --create-namespace \
  --version 26.3.1 \
  --set crds.enabled=true \
  --values <values-file>

Replace <values-file> with the file that holds your overrides. For the full procedure, see Upgrade the Redpanda Operator.

New features

  • RedpandaBrokerPool replaces NodePool for defining the pools of a Redpanda cluster, and requires an enterprise license (enterprise.licenseSecretRef in the operator chart). Without one, the operator makes no changes to the cluster’s resources or pools.

    Details

    To move a pool from a NodePool without rolling its brokers, install the experimental NodePool CRD (crds.experimental: true), run the operator with --enable-v2-nodepools, create a RedpandaBrokerPool with the same name, clusterRef, and spec, then annotate the NodePool with operator.redpanda.com/migrate-to-broker-pool: "true" or delete it. Do not downgrade the operator below this version once a RedpandaBrokerPool owns a pool that no NodePool defines.

  • Broker resources (beta): annotating a Redpanda resource with operator.redpanda.com/use-broker-cr moves its broker Pods from StatefulSets to per-broker Broker resources in place, with no restarts, and removing the annotation rolls back the same way. Migration and rollback progress shows in the BrokerMigration condition.

    Details

    The operator must run with --enable-broker, and the experimental Broker CRD must be installed (crds.experimental: true in the operator chart). Configuration changes made while the cluster uses Broker resources roll through the normal health-gated update flow, one broker at a time.

  • With Broker resources, a detect-disk-lost init container refuses to start Redpanda when a broker’s data volume is provably empty, so a wiped or replaced disk never re-registers the broker under a new node ID. The Broker is marked DiskLost, a replacement Broker takes its slot, and a DiskLostRemediated Warning event records the recovery.

  • When a broker refuses to start on a new major version because the cluster uses enterprise features without a license, Redpanda and StretchCluster resources now report it: LicenseValid is False with reason UpgradeRefused, status.license.violation is true, and a BrokerStartupRefused Warning event is emitted.

  • kubectl get -o wide for Redpanda and StretchCluster resources now shows the license type and expiration, whether enterprise features are in use without a valid license, and the enterprise features in use.

  • Pipeline resources have a new spec.updateStrategy field. Rolling starts a new Pod and waits for it to be ready before stopping an old one, so inputs that cannot replay, such as core NATS, no longer drop messages during restarts; Recreate stays the default.

    Details

    With Rolling, old and new Pods briefly overlap and may process some messages twice. Keep Recreate for inputs that need exclusive ownership, such as CDC replication slots.

  • Pipeline resources have a new spec.rpcPlugins field that lists Redpanda Connect dynamic (RPC) plugin manifest paths or globs. The manifests and plugin executables must already be in the Pod.

  • The Redpanda CRD gains spec.clusterSpec.service.internal.perPort. With enabled: true, the operator creates one Service per API and publishes a broker on each only while that API passes its health check, so clients avoid brokers that are restarting or, for Schema Registry, still catching up.

  • The operator can manage a Service’s EndpointSlices itself, so a broker leaves a Kafka port while that listener refuses connections, and leaves the Schema Registry ports while its registry is still catching up. Opt in by annotating a Cluster or Redpanda resource with operator.redpanda.com/enable-endpoint-steering=true. StretchClusters are not covered.

  • Rolling restarts now wait for each broker’s Schema Registry to catch up before rolling the next broker. The check is skipped on clusters without a Schema Registry listener, and --wait-for-schema-registry-sync=false disables it.

  • The operator chart can add a scrape-time label that identifies which Kubernetes cluster a metrics series came from, through monitoring.clusterLabel. It is off by default.

  • podTemplate now applies to every workload the operator chart renders, including the CRD installation Job and the post-upgrade migration Job. New deployment.podTemplate, jobs.crd.podTemplate, and jobs.migration.podTemplate values override each workload, and both Jobs now inherit the default pod securityContext (runAsUser: 65532).

  • The single-cluster operator supports --leader-election-lease-duration, --leader-election-renew-deadline, and --leader-election-retry-period for tuning leader election. The defaults are unchanged.

  • A new Prometheus gauge, redpanda_broker_phase{namespace, kind, name, phase}, counts the Broker resources of each cluster in each status.phase.

  • The StretchCluster controller watches its rendered resources in every member cluster, so drift in an owned resource triggers reconciliation immediately.

  • rpk k8s reports each command’s flags and usage line to rpk, so shell completion lists their flags and rpk --print-tree includes them.

  • Anonymous operator telemetry reports the use of Broker resources, and id_hash for clusters licensed through their resource, not only through the operator’s own license file.

  • The CRD reference documents every Console CRD field and the Broker resource, so kubectl explain describes them.

Improvements

  • Redpanda Connect support in the operator is now GA, including the Pipeline CRD and the Connect controller. The controller stays opt-in (--enable-connect / connectController.enabled) and requires an enterprise license with the Connect product.

  • The configurator and fs-validator script volumes on Redpanda and StretchCluster Pods are now named configurator and fs-validator, without the <fullname>- prefix, which keeps the configurator volume name within the Kubernetes 63-character limit for releases whose full name is 51 characters or longer. Brokers restart on upgrade, and a podTemplate override keyed on an old volume name must be re-keyed to the new name.

  • The configurator init container of StretchClusters now carries securityContext.runAsNonRoot: true and securityContext.allowPrivilegeEscalation: false, matching the chart and Redpanda Pods. StretchCluster brokers restart on upgrade.

  • The operator chart’s commonAnnotations value is renamed to connectAnnotations, and the --common-annotations flag to --connect-annotations. Rename it before upgrading: the chart’s values schema rejects unknown keys, so the old name fails helm upgrade.

  • An unresolvable external secret referenced from cluster config now names the affected property, and the Redpanda resource’s ConfigurationApplied condition and a Warning event carry the message.

  • StretchCluster lifecycle hooks now log an explicit TIMEOUT line when postStart.sh or preStop.sh is killed at its timeout.

  • StretchCluster sidecar and rack-awareness roles no longer grant permissions they never used. The operator owns these objects, so they narrow on upgrade with no action needed.

  • StretchCluster LoadBalancer Services now also carry a correctly spelled redpanda.com/type: loadbalancer label. The misspelled repdanda.com/type label is still set for existing selectors.

  • values.schema.json is now 3-5x smaller, which reduces the time every helm template, helm install, and helm upgrade spends compiling it. The schema accepts the same values, except that arrays nested inside a list or map of objects now accept null as well as a list.

  • The operator uses less CPU and memory by parsing the Redpanda chart’s default values once and caching the merged chart values for each distinct cluster spec.

  • With --leader-elect, the run command validates the leader election flags at startup and exits with an error that names the conflicting flags.

  • The default Redpanda Connect image for Pipeline resources is now 4.119.0, the default Redpanda Console image for Console resources is now v3.15.0, and the default Redpanda image tag (--redpanda-tag) is now v26.3.1.

Bug fixes

  • Operator upgrades now reach broker Pods. Before, the StatefulSet kept its old Pod template until the Redpanda resource or a NodePool changed. The first upgrade to this release rolls brokers once if their template changed since it was last applied.

    Details

    With image, statefulset.sideCars.image and statefulset.initContainerImage pinned and logging.usageStats.enabled: false, brokers roll only when a release changes the broker Pod spec.

  • Upgrading the operator no longer rolls every broker just because the embedded Redpanda chart’s version changed. The upgrade to this release rolls brokers once to drop the helm.sh/chart label.

  • A broker Pod that cannot start, for example one crash-looping after an upgrade or one whose image cannot be pulled, can now be fixed by changing the cluster resource. The operator replaces such a Pod once its Pod template changes, one Pod at a time, and never one whose broker other brokers still see running.

  • The operator recovers brokers wedged by a leader-restart ghost or a decommissioned broker’s bad_rejoin crashloop, on single-cluster and StretchCluster resources.

    Details

    The operator clears a leaked maintenance-mode flag on a dead broker whose address a live successor already owns. It can also delete the data directory PersistentVolumeClaim, if there is one, and the Pod of a broker whose on-disk identity collides with the cluster’s membership and that stays not ready past a threshold. It does this only when the cluster is healthy and no nodes are down. On the single-cluster operator, this deletion is disabled by default. To enable it, set --wipe-stale-disk-after to a positive duration.

  • The StatefulSet decommissioner no longer deletes the PersistentVolumeClaims of live brokers during a rolling restart. On Redpanda clusters, the operator-wide decommissioner deletes claims only when --auto-delete-pvcs is set.

  • When a Redpanda or StretchCluster StatefulSet scales down by more than one broker in a single change, the operator now waits for each removed Pod to be fully gone before it decommissions the next broker.

  • Brokers are no longer left in maintenance mode after a rolling restart when they take longer than the postStart hook timeout to start their Admin API.

  • V1 Cluster rolling restarts no longer delete a Pod undrained when Redpanda refuses to enable maintenance mode, and resume with the broker already in maintenance mode when the StatefulSet changes mid-drain.

  • Changing how a cluster’s internal listeners serve, such as enabling TLS or turning SASL on, no longer cuts the operator off from brokers it has yet to restart. TLS on the RPC listener still can’t be changed safely on a running cluster.

    Details

    A cluster already stuck on such a change when the operator is upgraded stays stuck: revert the change, wait for the cluster to settle, then apply it again. A Redpanda resource’s RPC listener follows tls.enabled unless listeners.rpc.tls.enabled is set, so set it to the listener’s current value before turning TLS on or off.

  • Clients in Redpanda and StretchCluster broker Pods now verify internal Admin API, Kafka API, and Schema Registry listeners that have a trustStore, and the sidecar and the post-install Job mount the trust stores. Brokers with a trustStore on one of these listeners restart once on upgrade.

  • The StretchCluster preStop lifecycle hook no longer stalls Pod termination or sends SIGTERM to a broker that is still draining, and bounds its Admin API calls.

  • StretchCluster startup and liveness probes now set timeoutSeconds: 5, so Admin API responses slower than 1s no longer restart a healthy broker. StretchCluster brokers restart once on upgrade.

  • StretchCluster clients no longer dial broker Pods that cannot be reached, including Pods on a node that has gone NotReady, and keep at least two admin hosts.

  • A StretchCluster reconcile that cannot fully read a member cluster now skips that member instead of acting on its partial state, and a persistently failing maintenance-mode, stale-disk-wipe, or decommission step no longer blocks cluster-config and license synchronization.

  • StretchCluster brokers now advertise their external address on external Kafka and HTTP Proxy listeners, and each pool uses its own external.addresses list. Pools that set external.domain or external.addresses advertise the new address after the next broker restart.

  • A StretchCluster listener with tls.enabled: true in a pool that disables TLS now gets its certificate. Before, the broker crashlooped.

  • StretchCluster LoadBalancer Services now publish the first advertised port of each external listener. Before, they published the listener port, but clients connect to the advertised port.

  • StretchCluster ClusterRole and ClusterRoleBinding names are no longer truncated to 63 characters, so pools no longer overwrite each other’s rules.

  • A retired node pool is now removed from status.nodePools on Redpanda and StretchCluster resources instead of lingering with its last replica counts.

  • NodePool brokers now inherit an unset image field from the Redpanda resource instead of the operator’s default image. NodePools that leave image unset roll to the Redpanda resource’s image on upgrade.

  • A Console or Pipeline that references a cluster through clusterRef now connects to the cluster’s NodePool brokers when the operator runs with --enable-v2-nodepools.

  • Redpanda clusters whose certificates set caEnabled: false are now reachable over TLS from the operator, Console, and Pipelines.

  • Resources that reference a cluster through cluster.clusterRef no longer fail to fetch the server root CA when the cluster’s certificate sets tls.certs.<cert>.issuerRef, and a CA bundle or a full chain works.

  • User, Role, and Group resources now reconcile on clusters with listeners.schemaRegistry.enabled=false.

  • SASL users from the users Secret are now created when the broker’s Admin API is not yet listening at the first sync. The sidecar retries with backoff.

  • The operator no longer sorts string-array cluster configuration values, so order-sensitive properties such as kafka_mtls_principal_mapping_rules keep the order given in the Redpanda resource.

  • When a broker rejects a cluster configuration write, the ConfigurationApplied condition and a Warning event carry the broker’s per-property reason instead of a bare Bad Request.

  • A ShadowLink whose link has been failed over no longer reports Synced=False and re-sends the refused update on every reconcile. The operator stops updating a link once failover starts and reports its state in status.state.

  • Setting post_install_job.enabled=false no longer prevents cluster configuration from being applied. Use config.extraClusterConfiguration or storage.tiered.credentialsSecretRef instead of environment injected through post_install_job.podTemplate.

  • Topic resources without spec.partitions or spec.replicationFactor no longer fail every reconcile after the topic is created.

  • An interrupted RedpandaRole rename no longer orphans the old role. The new status.pendingEffectiveRoleName field lets the next reconciliation retry the rename.

  • Pipelines with userRef now get their RPK_SASL_* environment variables from the referenced User instead of the cluster’s bootstrap user.

  • The Console configuration rendered for a Redpanda resource no longer enables SASL impersonation unless the legacy console.console block configured a login, so clusters without SASL no longer get a crash-looping Console.

  • With Broker resources, the PodDisruptionBudget now allows disruptions, so kubectl drain no longer hangs on nodes that run a broker. Broker resources also fix restarts after the first restart-requiring configuration change, and status for node pools scaled to zero.

  • The V1 Cluster controller no longer removes the cluster_id cluster configuration property on every configuration reassertion.

  • Deleting a Redpanda or StretchCluster no longer waits for the 3-minute periodic requeue before its finalizer is removed.

  • CRD installation no longer fails on Kubernetes versions that enforce CEL reserved-word escaping.

  • The multicluster operator now serves metrics over HTTPS by default, so the operator chart’s ServiceMonitor can scrape it, and it now reports telemetry.

  • The operator chart’s ServiceMonitors emit a lowercase scheme, so older prometheus-operator CRDs accept them.

  • The monitoring.rulesEnabled PrometheusRule and the Grafana dashboard now scope every expression to this operator install, so alerts no longer fire for other controllers or other operator installs. False-positive OperatorReconcileRunaway and OperatorWorkerPoolSaturated alerts no longer fire during healthy interval reconciliation, and a new OperatorReconcileChurn alert tracks non-interval work.

  • Objects rendered by the operator chart no longer carry annotations: null, so ArgoCD no longer reports them as permanently OutOfSync.

  • The post-upgrade migration Job no longer fails the upgrade when its ServiceAccount lacks permissions for optional resource types, and the field-manager migration now fully releases a resource whose only manager is a legacy one.

  • Nodepools after the first no longer inherit the first nodepool’s cluster.redpanda.com/nodepool label.

  • Official release images now embed the telemetry signing key, enabling operator telemetry.

  • The sidecar no longer writes malformed users-file lines, which can contain passwords, to its log.

  • The resources.memory.enable_memory_locking description no longer says to set it to true for production, and the cloud_storage_* field descriptions link to the Object Storage Properties page.

Deprecations

  • NodePool is deprecated in favor of RedpandaBrokerPool and will be removed in a future release.

  • The misspelled repdanda.com/type label on stretch-cluster LoadBalancer Services will be removed in the next major release.

Removed

  • nodePort on Redpanda external listeners has no effect. Use advertisedPorts instead.

  • The operator no longer migrates spec.clusterSpec.console in a Redpanda resource to a Console resource, and ignores the field. To deploy Redpanda Console, create a Console resource.

  • The deprecated legacy-decommission controller is removed. Use --additional-controllers=decommission, which all already selects.