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

v26.2.6
2026-11-11

Pipeline resources gain rolling updates and RPC plugin support, StretchCluster brokers now advertise their configured external addresses, and scaling down several brokers at once now decommissions each one before the next.

Yes

v26.2.5
2026-10-23

NodePool brokers inherit an unset image from the Redpanda resource, Console and Pipeline resources connect to NodePool brokers, and the operator and the default Redpanda image no longer trust the compromised eMudhra certificate authorities (CVE-2026-111943).

Yes

v26.2.4
2026-09-24

A rolling restart no longer deletes the PersistentVolumeClaims of live brokers, podTemplate applies to every workload the operator chart renders, and the commonAnnotations value is renamed to connectAnnotations.

Yes

v26.2.3
2026-09-03

Rolling restarts wait for Schema Registry to catch up on each broker, the multicluster metrics endpoint is served over HTTPS by default, and the operator’s alerts and Grafana dashboard are scoped to one operator install.

Yes

v26.2.2
2026-08-17

Recovery for brokers that crashloop with a retired identity, a fix for CRD installation on Kubernetes versions that enforce CEL reserved-word escaping, and correct cluster.redpanda.com/nodepool labels on every node pool.

Yes

v26.2.1
2026-07-28

Adds the beta Pipeline CRD for Redpanda Connect, per-broker probes that gate rolling restarts, and ShadowLink role and schema replication that is on by default, along with many StretchCluster and PVCUnbinder fixes.

Yes

v26.2.6 (2026-11-11)

Pipeline resources gain rolling updates and RPC plugin support, StretchCluster brokers now advertise their configured external addresses, and scaling down several brokers at once now decommissions each one before the next.

Before you upgrade
  • Pipeline resources that do not set spec.image move to Redpanda Connect 4.117.0.

  • Console resources that do not set an image move to Redpanda Console v3.14.0.

  • StretchCluster brokers restart once on upgrade to apply the new probe timeout.

  • Brokers managed by Redpanda resources restart once on upgrade to drop the helm.sh/chart label from their Pods.

  • StretchCluster pools that set external.domain or external.addresses advertise a new external address on Kafka and HTTP Proxy listeners after the next broker restart. Before you upgrade, create DNS records for the new names.

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

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 26.2.6 \
  --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

  • 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.

    Details

    For example, add a manifest in spec.configFiles with the plugin built into a custom spec.image, or copy in a static binary with spec.extraInitContainers.

Improvements

  • The default Redpanda Connect image for Pipeline resources is now 4.117.0. Pipeline resources that do not set spec.image move to it on upgrade.

  • The default Redpanda Console image for Console resources is now v3.14.0. Console resources that do not set an image move to it on upgrade.

Bug fixes

  • StretchCluster broker 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 to apply it, and a timeoutSeconds set through a RedpandaBrokerPool podTemplate still overrides it.

  • Upgrading the operator no longer restarts every broker just because the embedded Redpanda chart’s version changed. The upgrade to this release restarts brokers managed by Redpanda resources once to drop the helm.sh/chart label from their Pods.

  • 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. Previously, several broker Pods could be deleted at once with only the first decommissioned, so partitions with a majority of replicas on them lost quorum.

  • Topic resources without spec.partitions or spec.replicationFactor no longer fail every reconcile after the topic is created. The operator now leaves the topic’s current partition count and replication factor alone when those fields are unset.

  • StretchCluster brokers now advertise the same host as their LoadBalancer Service on external Kafka and HTTP Proxy listeners while external.enabled is true, and each pool now uses its own external.addresses list. Pools that set external.domain or external.addresses advertise the new address after the next broker restart, so create DNS records for those names before you upgrade.

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

  • Resources that reference a cluster through cluster.clusterRef, such as User, Topic, Schema, and Role, no longer fail with error fetching server root CA <namespace>/<cluster>-<cert>-root-certificate: server TLS certificate not found when the cluster’s certificate sets tls.certs.<cert>.issuerRef. For issuerRef and secretRef certificates alike, the operator now trusts every certificate in the issued Secret’s ca.crt when caEnabled is true, or its tls.crt otherwise, so a CA bundle or a full chain works.

  • A StretchCluster’s status.nodePools no longer keeps the entry of a removed node pool. An entry is removed only by a reconcile that observes every member cluster, so the pools of an unreachable member are kept.

  • A StretchCluster reconcile that cannot fully read a member cluster now skips that member instead of acting on its partial state, which could change a StatefulSet’s replica count without decommissioning the removed brokers. In flat network mode, resource rendering now waits until every member’s pools are fetched, instead of deleting the member’s per-pod Endpoints and EndpointSlices and breaking DNS to its brokers.

    Details

    While any member is skipped or its broker pools are not listed, the operator no longer scales a pool down without decommissioning. The ResourcesSynced condition names the members it waits for. Reads of a member’s Pods, ControllerRevisions, and StatefulSet selector now time out, so a member whose API server stops answering no longer stalls the reconcile.

  • 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, for Redpanda resources, the post-install Job now mount the trust stores so they can connect to the Admin API. Brokers with a trustStore on one of these listeners restart once on upgrade.

    Details

    If the listener’s certificate also has a CA (caEnabled), these clients verify it with a bundle of ca.crt and the trust store. A new server-ca-bundle init container writes the bundle when the Pod starts, and the sidecar rewrites it when ca.crt or the trust store changes, so clients pick up the change without a restart. A broker keeps verifying with the trust store alone until it restarts on the new Pod template.

v26.2.5 (2026-10-23)

NodePool brokers inherit an unset image from the Redpanda resource, Console and Pipeline resources connect to NodePool brokers, and the operator and the default Redpanda image no longer trust the compromised eMudhra certificate authorities (CVE-2026-111943).

Before you upgrade
  • Brokers managed by Redpanda and StretchCluster resources restart on upgrade, because the Pod template changes.

  • NodePools that leave image unset roll to the Redpanda resource’s image. To avoid a downgrade that Redpanda may refuse to start on, first set image.tag on the Redpanda resource or on the NodePool to the version the NodePool runs today.

  • If you override the configurator or fs-validator volumes or volumeMounts through podTemplate, re-key the override to the new names, configurator and fs-validator. An override keyed on an old name silently stops matching.

  • The default Redpanda image tag (--redpanda-tag) is now v26.2.6, and the default Redpanda Connect image for Pipeline resources is now 4.115.0.

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 26.2.5 \
  --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.

Improvements

  • StretchCluster lifecycle hooks now log a TIMEOUT line when postStart.sh or preStop.sh is killed at its timeout, matching the Helm chart. Previously, the hook was killed silently.

  • 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 StretchCluster brokers now sets securityContext.runAsNonRoot: true and securityContext.allowPrivilegeEscalation: false, matching the Redpanda Helm chart and Redpanda Pods. StretchCluster brokers restart on upgrade.

  • values.schema.json is now 3-5x smaller, so every helm template, helm install, and helm upgrade spends less time compiling it. Arrays nested inside a list or map of objects, such as podTemplate.spec.containers[].args, now accept null as well as a list.

  • The default Redpanda Connect image for Pipeline resources is now 4.115.0.

  • The default Redpanda image tag (--redpanda-tag) is now v26.2.6, whose image no longer trusts the compromised eMudhra certificate authorities (CVE-2026-111943).

Bug fixes

  • The operator no longer sorts string-array cluster configuration values before applying them. Order-sensitive properties such as kafka_mtls_principal_mapping_rules, sasl_kerberos_principal_mapping, and iceberg_default_catalog_namespace now keep the order given in the Redpanda resource.

  • The sidecar started by --watch-users now retries a user sync when the broker’s Admin API is not yet listening or returns an error. Previously, SASL users from the users secret were sometimes not created until the secret changed or the Pod restarted.

  • A Console that references a Redpanda cluster through clusterRef now connects to the cluster’s NodePool brokers. With --enable-v2-nodepools, Console no longer crash-loops after a NodePool migration drains the cluster’s own StatefulSet to statefulset.replicas: 0.

  • A Pipeline that references a Redpanda cluster through clusterRef now connects to the cluster’s NodePool brokers. With --enable-v2-nodepools, a Pipeline no longer loses its seed brokers after a NodePool migration drains the cluster’s own StatefulSet to statefulset.replicas: 0.

  • NodePool brokers now inherit any unset image field from the Redpanda resource instead of the operator’s default Redpanda image, which produced a mixed-version cluster when the Redpanda resource pinned image.tag. On upgrade, NodePools that leave image unset roll to the Redpanda resource’s image, so first set image.tag to the version the NodePool runs today.

  • Brokers managed by the Redpanda and StretchCluster resources are no longer left in maintenance mode after a rolling restart when their Admin API takes longer than the postStart hook timeout to start.

  • The operator is now built with Go 1.26.10, which rejects certificates chaining to the compromised eMudhra intermediate certificate authorities (CVE-2026-111943) and fixes several HTTP/2 denial-of-service vulnerabilities.

  • golang.org/x/net is updated to v0.60.0, golang.org/x/crypto to v0.57.0, and OpenTelemetry to v1.46.0 (logs v0.22.0) to address reported vulnerabilities, including a busy-spin in the OpenTelemetry log batch processor (CVE-2026-81872).

v26.2.4 (2026-09-24)

A rolling restart no longer deletes the PersistentVolumeClaims of live brokers, podTemplate applies to every workload the operator chart renders, and the commonAnnotations value is renamed to connectAnnotations.

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

  • The CRD installation Job and the post-upgrade migration Job now inherit the default Pod securityContext (runAsUser: 65532) from podTemplate.

  • Environment variables injected through post_install_job.podTemplate no longer feed cluster configuration. Use config.extraClusterConfiguration or storage.tiered.credentialsSecretRef instead.

  • On Redpanda (v2) clusters, the operator-wide decommissioner deletes PersistentVolumeClaims only when --auto-delete-pvcs is set. Pass the flag if you rely on the decommissioner to remove the claims that a scale-down leaves behind.

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 26.2.4 \
  --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

  • podTemplate now applies to every workload the chart renders: the operator Deployment, the CRD installation Job, and the post-upgrade migration Job. Both Jobs now inherit its default Pod securityContext (runAsUser: 65532), which previously reached only the Deployment.

    Details

    To override one workload, set deployment.podTemplate, jobs.crd.podTemplate, or jobs.migration.podTemplate. Each one layers on top of podTemplate, which makes the container securityContext of the hook Jobs configurable. Pod templates are merged by name, so to override one container, give the name of that container in the override.

Improvements

  • Breaking change for Redpanda Connect pipelines, which are still in beta: the operator chart’s commonAnnotations value is renamed to connectAnnotations, and the operator’s --common-annotations flag to --connect-annotations. Rename the value before you upgrade, because helm upgrade fails on the old name.

Bug fixes

  • Setting post_install_job.enabled=false no longer prevents cluster configuration from being applied. Environment variables injected through post_install_job.podTemplate no longer feed cluster configuration, so use config.extraClusterConfiguration or storage.tiered.credentialsSecretRef instead.

  • In the multicluster operator, a persistently failing maintenance-mode, stale-disk-wipe, or decommission step no longer blocks a StretchCluster’s cluster configuration and license synchronization. Previously, ConfigurationApplied could report True indefinitely while spec changes were never applied.

  • StretchCluster admin, Kafka, and Schema Registry clients no longer connect to broker Pods on a node that has gone NotReady. Previously, connection timeouts to those unreachable Pods stalled decommission and pool cleanup.

  • StretchCluster clients keep at least two admin hosts when endpoint filtering leaves a single broker, instead of failing every request against that one broker.

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

  • Objects rendered by the operator chart no longer carry annotations: null when no annotations are configured. Previously, Argo CD kept the ServiceMonitor and every other chart object permanently OutOfSync.

  • The Console configuration rendered for a Redpanda resource no longer contains kafka.sasl.enabled: true and kafka.sasl.impersonateUser: true unless the legacy console.console block configures a login. Previously, Console 3.9 crash-looped on clusters without SASL.

  • The StatefulSet decommissioner no longer deletes the PersistentVolumeClaims of live brokers during a rolling restart. On Redpanda (v2) clusters, the operator-wide decommissioner now deletes claims only when --auto-delete-pvcs is set.

    Details

    Previously, updating a StatefulSet could delete the claims of Pods that had not yet rolled. Those Pods came back onto empty volumes and rejoined the cluster as new brokers. The decommissioner started by --additional-controllers=decommission (or all) also deleted claims by default. If you rely on it to remove the claims that a scale-down leaves behind, pass --auto-delete-pvcs.

v26.2.3 (2026-09-03)

Rolling restarts wait for Schema Registry to catch up on each broker, the multicluster metrics endpoint is served over HTTPS by default, and the operator’s alerts and Grafana dashboard are scoped to one operator install.

Before you upgrade
  • The multicluster command serves its metrics endpoint over HTTPS by default. If you overrode the ServiceMonitor to scheme: http, remove that override, or pass --metrics-secure=false to keep the plain-HTTP endpoint.

  • Rolling restarts wait for each broker’s Schema Registry store to catch up before rolling the next broker. To disable this behavior, pass --wait-for-schema-registry-sync=false.

  • The multicluster operator starts sending telemetry reports. Previously, it dropped every report.

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 26.2.3 \
  --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

  • The operator chart can add a label to operator metrics that identifies which Kubernetes cluster each series came from. Set monitoring.clusterLabel.enabled to true (off by default) when metrics from more than one Kubernetes cluster land in the same Prometheus-compatible backend.

    Details

    In multicluster mode, the label value defaults to multicluster.name. Otherwise, set monitoring.clusterLabel.value.

  • Rolling restarts now wait for each broker’s Schema Registry store to catch up on _schemas before rolling the next broker, so a cluster keeps a consistent Schema Registry endpoint during an upgrade. This applies to Cluster, Redpanda, and StretchCluster resources, and is skipped on clusters without a Schema Registry listener. To disable it, pass --wait-for-schema-registry-sync=false.

  • The single-cluster operator supports --leader-election-lease-duration, --leader-election-renew-deadline, and --leader-election-retry-period for tuning leader election. The existing defaults of 15s, 10s, and 2s are unchanged.

  • Telemetry now reports id_hash for clusters licensed through their custom resource, not only for operators given their own license file. A new clusterLicenses field carries the distinct license checksums found across managed clusters and the count of validly licensed clusters.

Bug fixes

  • The sidecar no longer writes malformed users-file lines, which can contain passwords, to its log. When the sidecar manages the superusers cluster property, it now sets the property once per sync pass with the users from every users file, instead of once per file.

  • The multicluster command now serves its metrics endpoint over HTTPS by default, so the operator chart’s ServiceMonitor can scrape a multicluster deployment. If you overrode the ServiceMonitor to scheme: http, remove that override, or pass --metrics-secure=false to keep the plain-HTTP endpoint.

  • An unreadable broker list no longer stops the whole reconcile pass. Previously, a transient read failure against a broker whose Pod was already gone also stopped decommissioning and cluster configuration sync, so that broker was never removed.

  • The StretchCluster admin, Kafka, and Schema Registry clients no longer try brokers whose Pod cannot be reached, such as a terminating or finished Pod. If no Pod looks reachable, the clients use the full declared broker list as before.

  • The monitoring.rulesEnabled alerts and the shipped Grafana dashboard now apply only to this operator install, by job and namespace. Previously, the alerts could fire for other controllers in the same Prometheus, such as Karpenter, Flux, cert-manager, or a second operator install. The dashboard gains Namespace and Job pickers, plus one for the monitoring.clusterLabel label.

  • The OperatorReconcileRunaway and OperatorWorkerPoolSaturated alerts no longer fire during healthy interval reconciliation. A new OperatorReconcileChurn alert tracks non-interval work, and OperatorWorkerPoolSaturated now measures sustained workqueue delay instead of worker utilization.

  • User, Role, and Group resources now reconcile on clusters with listeners.schemaRegistry.enabled=false. Previously, a disabled Schema Registry listener blocked all ACL syncs.

  • The multicluster operator now reports telemetry. Previously, it dropped every report.

v26.2.2 (2026-08-17)

Recovery for brokers that crashloop with a retired identity, a fix for CRD installation on Kubernetes versions that enforce CEL reserved-word escaping, and correct cluster.redpanda.com/nodepool labels on every node pool.

Before you upgrade
  • The operator needs a new permission: get on pods/log. The generated roles include it.

  • Operator telemetry is enabled in the official release images. In earlier images, including v26.2.1, the telemetry client was disabled.

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 26.2.2 \
  --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.

Bug fixes

  • The operator now recovers a broker that crashloops because it starts with a retired identity that the controller refuses (bad_rejoin), on single-cluster and StretchCluster resources, with or without persistent storage. The operator needs a new permission for this: get on pods/log, which the generated roles include.

    Details

    The operator clears a leaked maintenance-mode flag on a dead broker whose address a live successor already owns. This does not depend on --wipe-stale-disk-after. The operator 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, so that the broker restarts with a new identity. 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. Zero or a negative value disables it. The StretchCluster operator keeps its existing default of 5m.

  • On StretchCluster resources, the preStop lifecycle hook no longer stalls Pod termination or stops a broker in the middle of a drain. The hook skips the drain for a broker that is no longer a cluster member, and every admin API call in the hook has a timeout.

  • CRD installation no longer fails on Kubernetes versions that enforce CEL reserved-word escaping (observed on 1.31). Previously, a fresh install of operator v26.2.1 with crds.enabled=true could not complete on those versions.

  • Each node pool’s Pods now carry their own cluster.redpanda.com/nodepool label. Previously, every pool after the first got the first pool’s label.

  • The operator chart’s ServiceMonitor and the StretchCluster per-pool ServiceMonitor resources now use lowercase scheme: http or scheme: https. Older prometheus-operator CRDs reject the uppercase values.

  • The post-upgrade migration job no longer fails the whole upgrade when its ServiceAccount lacks permissions for optional resource types, such as Gateway API tlsroutes or httproutes on clusters that manage RBAC manually (rbac.create=false). The job logs a warning, skips the resource, and checks it again on the next upgrade.

  • Operator telemetry now works in the official release images. In earlier images, including v26.2.1, the telemetry client was disabled.

v26.2.1 (2026-07-28)

Adds the beta Pipeline CRD for Redpanda Connect, per-broker probes that gate rolling restarts, and ShadowLink role and schema replication that is on by default, along with many StretchCluster and PVCUnbinder fixes.

Before you upgrade
  • Brokers restart once, in a rolling restart, when you upgrade the operator or chart.

  • If you pin --configurator-tag or sideCars.image.tag to an older operator release, bump the pin to this release or later.

  • Existing ShadowLinks start replicating all RBAC roles and the source cluster’s schemas topic on the first reconcile after upgrade. To keep the previous behavior, set spec.roleSyncOptions.enabled: false and spec.schemaRegistrySyncOptions.enabled: false before you upgrade.

  • A User whose spec.authentication.type was edited but never took effect gets the change applied on the first reconcile, and the credential under the previous mechanism stops working. Migrate clients to the requested mechanism before you upgrade.

  • The SASL mechanism fields on the Redpanda and StretchCluster CRDs accept only SCRAM-SHA-256 or SCRAM-SHA-512. The next update of an object with any other value, including mechanism: "", is rejected.

  • A clusterRef.group value that contains a version suffix, such as cluster.redpanda.com/v1alpha2, is rejected by CRD validation.

  • The nodeWatcher value for --additional-controllers is no longer recognized. Remove it and rely on --unbind-pvcs-after.

  • --additional-controllers=decommission and all now run the NodePool-aware StatefulSetDecommissioner. No configuration change is required.

  • Existing Topics keep reconciling every 3s because spec.interval: "3s" is stored on them. Topics with no stored spec.interval use the operator default of 30s.

  • A Console whose clusterSource.clusterRef sets namespace now resolves its cluster in that namespace instead of the Console’s own namespace.

  • For a StretchCluster, the operator now deletes the Pod and PVC of a broker whose stale disk collides with the cluster’s view after the Pod is not-Ready for 5m, when the cluster is healthy and no nodes are down. Set --wipe-stale-disk-after to a negative value to disable this.

  • On multicluster installs, the operator chart grants the operator ServiceAccount nodes: get/list/watch. The PVCUnbinder’s ClusterRole also grants nodes list and storageclasses get, so update RBAC that you manage outside the chart.

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 26.2.1 \
  --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

  • Console can be exposed through Gateway API controllers as an alternative to Ingress. The Console CRD supports HTTPRoute.

  • The operator uses the per-broker pre-restart and post-restart probes in Redpanda 25.1+ to gate rolling restarts of Redpanda and StretchCluster resources: it restarts a broker only when the probe reports no risk of data loss or unavailability (a topic with a replication factor of 1 going offline is accepted), and waits for that broker to finish recovery before it restarts the next one. Clusters older than 25.1 keep the cluster-wide health check, and --post-restart-caught-up-percent (default 100) sets how much recovery is required.

  • clusterRef accepts an optional namespace field on ShadowLink, Topic, User, Role, Schema, and Group resources, so a resource can reference a cluster in another namespace. A ShadowLink between two namespaces on the same Kubernetes cluster is intended mainly for development and test clusters.

  • The operator can report optional anonymous telemetry about cluster shape, such as cluster and broker counts, Redpanda versions in use, and feature adoption, and licensed clusters also report the enterprise license checksum. Disable it with the telemetry.enabled chart value or --disable-telemetry. Telemetry does not run in the published v26.2.1 images. It works from v26.2.2.

  • The operator Helm chart has a top-level priorityClassName value. Set it to a higher-priority PriorityClass so the operator Pod is less likely to be preempted when the cluster is under resource pressure.

  • The rpk-k8s CLI is published as a managed rpk plugin with each operator release. Run it with rpk k8s, and rpk downloads, verifies, and caches it on first use.

  • A Broker custom resource and controller manage the Pod and PVC lifecycle of individual brokers. This is experimental and runs only when you set --enable-broker.

  • Kubernetes versions 1.33.x through 1.36.x are supported.

  • The operator automatically clears maintenance mode on a broker that is stuck in it, for example because its Pod cannot reschedule from a cordoned or lost node, which unblocks auto-decommission. This applies to Redpanda and StretchCluster resources when the cluster reports the broker as not alive and the Pod is not-Ready for 30m by default, and --clear-maintenance-mode-after changes the threshold.

  • For a StretchCluster, the operator can recover a broker that crash-loops after a region outage because it boots from a stale disk: it deletes the Pod and its PVC so that the broker reschedules onto a fresh disk. This runs only when the broker’s on-disk identity collides with the cluster’s view, the cluster is healthy, and the Pod is not-Ready for 5m by default. Set --wipe-stale-disk-after to change the threshold, or to a negative value to disable it.

  • ShadowLink resources expose spec.clientOptions to tune the Kafka client that the shadow cluster uses to fetch from the source cluster. Unset fields keep the broker default, so existing ShadowLinks are unaffected.

  • ShadowLink replicates RBAC roles and schemas from the source cluster by default, controlled by spec.roleSyncOptions and spec.schemaRegistrySyncOptions. Existing ShadowLinks start replicating all roles and the source cluster’s schemas topic on the first reconcile after upgrade.

    Details

    Replicated roles may include privileged or superuser roles. To keep the previous behavior, set spec.roleSyncOptions.enabled: false and spec.schemaRegistrySyncOptions.enabled: false on existing ShadowLinks before you upgrade the operator. Schemas can also be replicated from an external schema registry, for example a Confluent Schema Registry, with spec.schemaRegistrySyncOptions.shadowSchemaRegistryAPI.

  • The Pipeline CRD and Connect controller (beta) run Redpanda Connect pipelines declaratively, with support for kubectl scale and HorizontalPodAutoscaler, and roll pipeline Pods when the configuration or a referenced Secret or ConfigMap changes. This is an enterprise feature and is disabled by default.

    Details

    To opt in, set connectController.enabled: true in the operator Helm chart values, or pass --enable-connect. Configure the enterprise license for the operator with enterprise.licenseSecretRef in the chart values.

Improvements

  • --additional-controllers=decommission now runs the NodePool-aware StatefulSetDecommissioner for Redpanda clusters, which is correct for clusters split across multiple NodePools. Operators that already pass decommission or all switch to it on upgrade with no configuration change.

  • The Console controller now honors the namespace field of clusterSource.clusterRef. Previously, it always resolved the cluster in the Console’s own namespace, so a Console could fail to find its cluster or bind to a same-named cluster in the wrong namespace.

  • Sync intervals for the Topic, User, Group, Schema, Role, and ShadowLink controllers are now configurable with flags such as --topic-sync-interval or with the controllers.<resource>.interval Helm chart values. The Topic CRD no longer sets a default interval of 3s, so new Topics use the operator default of 30s, but existing Topics keep reconciling every 3s until you clear the stored value.

    Details

    Topic defaults to 30s. User, Group, Schema, Role, and ShadowLink default to 5m. A spec.interval set on a Topic takes precedence over the operator default. To make an existing Topic adopt the operator default, remove its stored value: kubectl patch topic <name> --type=json -p='[{"op":"remove","path":"/spec/interval"}]'

  • The SASL mechanism fields on the Redpanda and StretchCluster CRDs now accept only SCRAM-SHA-256 or SCRAM-SHA-512, so an invalid value is rejected at apply time. Existing objects are not re-checked on upgrade, but the next update of an object with any other value, including mechanism: "", is rejected, so correct the value or omit the field.

Bug fixes

  • A Redpanda resource that renders TLSRoute objects through the chart’s Gateway API support no longer causes a no kind "TLSRouteList" / "ListOptions" is registered for version "gateway.networking.k8s.io/v1" error on every reconcile.

  • A single reconcile no longer deletes more than one broker Pod during a rolling restart, for Redpanda and StretchCluster resources. Previously, a transient Admin API error or an address-format mismatch could make the operator treat every Pod as an orphan and delete them all at once.

  • The broker decommissioner (--additional-controllers=decommission) now builds the correct broker hostname when fullnameOverride is set. Previously, its health check failed with no such host and decommission and PVC cleanup never completed.

  • The broker decommissioner now handles clusters split across multiple NodePools: it compares the total broker count against the sum of every pool’s desired replicas and acts only on the Pods of each pool’s own StatefulSet.

  • The PVCUnbinder no longer acts on several Pending broker Pods at once during a cluster-wide disruption, which could delete every broker’s PVCs in parallel. It now defers when stuck Pods span more than one node, when another unbind is in progress, or when the parent resource has the operator.redpanda.com/pause-pvc-unbinder=true annotation.

    Details

    A new --broker-pod-node-unavailable-toleration flag adds long NotReady and Unreachable tolerations to broker Pods, so brief node unreachability does not lead to an unbind. --allow-pv-rebinding is deprecated because the new checks make it unnecessary and unsafe.

  • The Topic controller no longer issues a no-op IncrementalAlterConfigs request on every reconcile. It writes only the configuration keys that differ from the live broker configuration.

  • The Topic, User, Group, Schema, Role, and ShadowLink controllers no longer reconcile every resource on each status-only update of the referenced cluster, so configured sync intervals such as the Topic spec.interval take effect.

  • clusterSpec.statefulset.extraVolumeMounts, livenessProbe, and startupProbe now apply to the redpanda container. Previously, they applied only to the sidecar container.

  • The rpk configuration inside broker Pods no longer goes stale after brokers join or leave the cluster, for both Cluster and Redpanda resources. Brokers restart once on upgrade, and if you pin --configurator-tag or sideCars.image.tag to an older operator release, you must bump the pin to this release or later.

  • clusterSpec.statefulset overrides such as probes and extraVolumeMounts are no longer dropped when podTemplate.spec.containers or podTemplate.spec.initContainers contains duplicate entries with the same name.

  • Flags set in additionalCmdFlags now override the defaults that the chart sets, such as --enable-console. Previously, they were silently dropped.

  • Deleting a User no longer gets stuck on clusters without an enterprise license. Previously, Schema Registry ACL cleanup failed with "Invalid license: not present" and the finalizer was never removed.

  • A StretchCluster peer whose StretchCluster resource has been removed is no longer reported as SpecSynced=ClusterUnreachable, and no longer blocks the cluster from reaching Stable.

  • A StretchCluster with SASL authentication disabled now sets the BootstrapUserSynced condition to True. Previously, the condition stayed Unknown, which kept an otherwise healthy cluster from reaching Stable.

  • Operator logs for StretchCluster rolling restarts and StatefulSet changes now include the cluster name instead of a blank identifier.

  • The Redpanda and StretchCluster reconcilers no longer stall recovery when no broker is Ready, for example during a full-region outage. The operator now still clears stuck maintenance mode, decommissions brokers, and, for a StretchCluster, wipes stale disks.

  • A StretchCluster with rackAwareness.enabled: true can now deploy from a default chart install. On multicluster installs, the operator chart now grants the operator ServiceAccount nodes: get/list/watch.

  • Setting tls.enabled: false on a RedpandaBrokerPool now also disables TLS on the default external listeners, so brokers no longer crash-loop with "Could not read trust file". Also, a clusterRef.group value that contains a version suffix, such as cluster.redpanda.com/v1alpha2, is now rejected by CRD validation instead of silently never binding.

  • The User CRD now accepts spec.authentication.password.value on its own, as documented. Previously, password.valueFrom was always required.

  • The PVCUnbinder no longer deadlocks a cluster that uses WaitForFirstConsumer volume binding when a Pending Pod holds a local volume on a node that is unavailable or already occupied. Previously, a fresh cluster in this state never bootstrapped.

    Details

    The unbinder’s ClusterRole now also grants nodes list and storageclasses get. If RBAC that you manage lacks them, the unbinder keeps deferring. To turn off only this behavior, set --disable-pvc-rebinding-gate-exemption, or statefulset.sideCars.pvcUnbinder.disableStuckClaimExemption in the redpanda chart. When two brokers are affected on two nodes, you must still delete the PVC manually.

  • The operator now clears maintenance mode on a dead broker ID that was replaced by a live broker at the same address, which happens when a Pod loses its data directory and rejoins the cluster. This requires a stable advertised DNS address, so brokers that advertise bare Pod IPs are not matched.

  • Applying the same StretchCluster manifest to every member cluster at once, for example with a GitOps controller, no longer stores different bootstrap-user passwords across clusters. Previously, this permanently halted reconciliation with Cluster has no desired brokers.

  • StretchCluster now honors spec.auth.sasl.bootstrapUser.secretKeyRef and replicates the password to every member cluster. Previously, the field was ignored and the operator always used its own <name>-bootstrap-user Secret.

  • A StretchCluster whose bootstrapUser.secretKeyRef points at a Secret with no data under the referenced key now reports an error on the BootstrapUserSynced condition. Previously, the operator generated a new password and wrote it to only some member clusters.

  • Adding a bootstrapUser.secretKeyRef to a StretchCluster that was bootstrapped with the operator-managed <name>-bootstrap-user Secret no longer breaks operator authentication. The operator keeps the original password and copies it to the referenced location, and reports an error if that location already holds a different password.

  • SASL authentication now works on StretchClusters that use the default SCRAM mechanism, so User resources and ACL sync no longer fail with SASL_AUTHENTICATION_FAILED. Changing auth.sasl.mechanism on a cluster that is already bootstrapped still causes the same failure.

  • StretchCluster lifecycle hooks now clear maintenance mode left behind by a previous incarnation of the Pod that re-registered under a different node ID.

  • The StretchCluster controller no longer renders broker StatefulSets and ConfigMaps when it cannot observe a peer cluster’s broker pools. Previously, a broker in a newly added RedpandaBrokerPool could start with only itself in seed_servers and form a separate cluster.

  • Changing spec.authentication.type on an existing User now rewrites the user’s SCRAM credential under the new mechanism. Previously, nothing was written and the User still reported Synced=True.

    Details

    For a User whose type was edited but never took effect, the first reconcile after upgrade applies the change, and the credential under the previous mechanism stops working. Migrate clients first. Deleting a User now removes the mechanisms that the cluster actually holds. On brokers too old to support the Kafka SCRAM describe API, a type change is still not detected.

Deprecations

  • The --allow-pv-rebinding flag is deprecated because the new PVCUnbinder checks make it unnecessary and unsafe.

  • The old broker DecommissionReconciler is deprecated and will be removed in a future release. It is available only with --additional-controllers=legacy-decommission, which all does not include.

Removed

  • The nodeWatcher controller is removed, and the nodeWatcher value for --additional-controllers is no longer recognized. Remove it and rely on the PVCUnbinder (--unbind-pvcs-after) for PVC remediation after node loss.