Streaming

Beta

Migrate Node Pools

You can migrate Redpanda brokers to a new Kubernetes node pool using the NodePool custom resource (CR). This approach performs a blue/green migration: it creates new brokers (green) on the target node pool, replicates data, then decommissions old brokers (blue). The Redpanda Operator manages the decommission process automatically.

After reading this page, you will be able to:

Perform a blue/green node pool migration using the NodePool CRD
Verify cluster health and data integrity before and after migration

Use this approach when you need to upgrade Redpanda versions, change node instance types, perform vertical scaling, or upgrade Kubernetes nodes. For the previous manual method that does not use the NodePool CRD, see Migrate Node Pools Manually.

The NodePool CRD is a beta feature. You must enable experimental CRDs and the --enable-v2-nodepools operator flag to use it.

How it works

The NodePool migration follows a blue/green pattern:

You verify that the existing (blue) NodePool is healthy.
You create a new (green) NodePool with the target Redpanda version, optionally targeting specific Kubernetes nodes with nodeSelector and tolerations.
New green brokers join the cluster and begin replicating data. The cluster temporarily runs 2x the usual number of brokers.
You delete the old (blue) NodePool. The operator decommissions blue brokers one at a time, draining partitions to the green brokers automatically.
Only green brokers remain. You clean up the old Kubernetes node pool.

During migration, both NodePools serve traffic. No manual decommission commands are needed — deleting the old NodePool triggers the operator to handle everything.

The NodePool’s spec.image.tag field controls which Redpanda version the brokers run. NodePools do not inherit spec.clusterSpec.image.tag from the Redpanda CR. You must set the image tag explicitly on each NodePool.

flowchart TB A(("Start")) --> B["Verify blue NodePool is healthy"] B --> C["Create green NodePool with target version"] C --> D["Green brokers join cluster (2x brokers)"] D --> E["Wait for all brokers to be healthy"] E --> F["Delete blue NodePool"] F --> G["Operator decommissions blue brokers one by one"] G --> H["Verify only green brokers remain"] H --> I["Delete old Kubernetes node pool"] I --> J(("Complete")) classDef userAction stroke:#374D7C, fill:#E2EBFF, font-weight:bold, rx:5, ry:5; classDef systemAction fill:#F6FBF6, stroke:#25855a, stroke-width:2px, color:#20293c, rx:5, ry:5; class A,J systemAction; class B,C,E,F,H,I userAction; class D,G systemAction; click B "#verify-blue" click C "#create-green" click E "#wait-healthy" click F "#delete-blue" click H "#verify-green" click I "#cleanup"

Prerequisites

A Redpanda cluster deployed with the Redpanda Operator using the NodePool CRD. See Node Pools for setup instructions, including how to migrate an existing cluster to NodePools.
The Redpanda Operator deployed with experimental CRDs and the --enable-v2-nodepools flag enabled.
The Redpanda CR must have statefulset.replicas set to 0 so that NodePool CRs control the broker count.
An existing (blue) NodePool CR running your cluster.
If targeting specific Kubernetes nodes: a new Kubernetes node pool provisioned in your hosting platform (GKE, EKS, AKS, or self-managed) with nodes that meet Redpanda requirements. See Control node placement with nodeSelector and tolerations for setup.
A staging environment in which to test the migration before performing it in production.
Confirmation that the target Redpanda version is compatible with your Operator version. See Kubernetes Compatibility.

Pre-migration considerations

Before migrating, consider the following:

Cost: During migration, the cluster runs 2x the usual number of brokers. Plan for temporary increased infrastructure costs.
Data volume: Decommissioning time depends on the amount of data to replicate. Large datasets may take hours per broker.
Replication factor: Your broker count should exceed your highest topic replication factor to maintain availability during decommissioning.
Partition count: Total partition count is generally determined by the overall CPU core count. Verify the new nodes can handle the partition load.
Image tag: You must set spec.image.tag on the green NodePool explicitly. NodePools do not inherit the image tag from the Redpanda CR.

Verify the blue NodePool

Before starting the migration, verify that the existing blue NodePool is running and healthy. This example assumes you are upgrading from Redpanda v26.1.1 to v26.1.3.

Check the NodePool status:

kubectl get nodepool -n <namespace>

Expected output

NAME   CLUSTER    REPLICAS   READY   AGE
blue   redpanda   3          3       24h

Verify the NodePool conditions:

kubectl describe nodepool blue -n <namespace>

Confirm that the conditions show Bound=True, Deployed=True, and Stable=True.

Verify the Pods are running:

kubectl get pods -n <namespace> -l app.kubernetes.io/name=redpanda -o wide

Expected output

NAME              READY   STATUS    RESTARTS   AGE   IP            NODE
redpanda-blue-0   2/2     Running   0          24h   10.244.1.5    worker-1
redpanda-blue-1   2/2     Running   0          24h   10.244.2.5    worker-2
redpanda-blue-2   2/2     Running   0          24h   10.244.3.5    worker-3

Check cluster health:

kubectl exec -n <namespace> redpanda-blue-0 -c redpanda -- rpk cluster health

Expected output

CLUSTER HEALTH OVERVIEW
=======================
Healthy:                          true
Unhealthy reasons:                []
Controller ID:                    0
All nodes:                        [0 1 2]
Nodes down:                       []
Leaderless partitions (0):        []
Under-replicated partitions (0):  []

List the brokers and confirm their current version:

kubectl exec -n <namespace> redpanda-blue-0 -c redpanda -- rpk redpanda admin brokers list

Expected output

NODE-ID  NUM-CORES  MEMBERSHIP-STATUS  IS-ALIVE  BROKER-VERSION
0        1          active             true      v26.1.1
1        1          active             true      v26.1.1
2        1          active             true      v26.1.1

All three brokers should show membership-status: active, is-alive: true, and the current Redpanda version (v26.1.1 in this example).

Create the green NodePool

Create a new (green) NodePool with the target Redpanda version. In this example, you are upgrading from v26.1.1 to v26.1.3:

nodepool-green.yaml

apiVersion: cluster.redpanda.com/v1alpha2
kind: NodePool
metadata:
  name: green
  namespace: <namespace>
spec:
  clusterRef:
    name: redpanda
  replicas: 3
  image:
    tag: v26.1.3 (1)
  nodeSelector:
    nodetype: redpanda-green (2)
  tolerations:
    - key: redpanda-green (3)
      operator: Equal
      value: "true"
      effect: NoSchedule

1	The target Redpanda version. This must be set explicitly — NodePools do not inherit the image tag from the Redpanda CR.
2	Optional. Constrains green brokers to Kubernetes nodes labeled `nodetype: redpanda-green`. Omit if you don’t need node placement control.
3	Optional. Allows scheduling on nodes tainted with `redpanda-green:NoSchedule`. Omit if you don’t use taints.

If you don’t need to pin brokers to specific Kubernetes nodes, you can omit nodeSelector and tolerations. The green brokers will be scheduled on any available nodes.

Apply the NodePool:

kubectl apply -f nodepool-green.yaml

Wait for all brokers to stabilize

The cluster temporarily runs 2x the usual broker count (6 brokers in this example). Wait for the green Pods to become ready:

kubectl get pods -n <namespace> -l app.kubernetes.io/name=redpanda -o wide

Expected output (6 Pods)

NAME               READY   STATUS    RESTARTS   AGE     IP            NODE
redpanda-blue-0    2/2     Running   0          24h     10.244.1.5    worker-1
redpanda-blue-1    2/2     Running   0          24h     10.244.2.5    worker-2
redpanda-blue-2    2/2     Running   0          24h     10.244.3.5    worker-3
redpanda-green-0   2/2     Running   0          2m      10.244.4.5    worker-4
redpanda-green-1   2/2     Running   0          2m      10.244.5.5    worker-5
redpanda-green-2   2/2     Running   0          2m      10.244.6.5    worker-6

Wait for the Redpanda cluster to report as healthy with all 6 brokers:

kubectl exec -n <namespace> redpanda-blue-0 -c redpanda -- rpk cluster health

Expected output

CLUSTER HEALTH OVERVIEW
=======================
Healthy:                          true
Unhealthy reasons:                []
Controller ID:                    0
All nodes:                        [0 1 2 3 4 5]
Nodes down:                       []
Leaderless partitions (0):        []
Under-replicated partitions (0):  []

Verify that all 6 brokers are active and confirm the version split:

kubectl exec -n <namespace> redpanda-blue-0 -c redpanda -- rpk redpanda admin brokers list

Expected output

NODE-ID  NUM-CORES  MEMBERSHIP-STATUS  IS-ALIVE  BROKER-VERSION
0        1          active             true      v26.1.1
1        1          active             true      v26.1.1
2        1          active             true      v26.1.1
3        1          active             true      v26.1.3
4        1          active             true      v26.1.3
5        1          active             true      v26.1.3

Do not proceed until the cluster reports as Healthy: true with zero leaderless and zero under-replicated partitions. All 6 brokers must show membership-status: active and is-alive: true.

Delete the blue NodePool

Delete the old (blue) NodePool. The operator decommissions each blue broker one at a time, draining partitions to the green brokers before removing each one. No manual decommission commands are needed.

kubectl delete nodepool blue -n <namespace>

Monitor the decommission progress:

kubectl exec -n <namespace> redpanda-green-0 -c redpanda -- rpk redpanda admin brokers list

You will see blue brokers transition from active to draining as the operator decommissions them:

Example output (decommission in progress)

NODE-ID  NUM-CORES  MEMBERSHIP-STATUS  IS-ALIVE  BROKER-VERSION
0        1          draining           true      v26.1.1
1        1          active             true      v26.1.1
2        1          active             true      v26.1.1
3        1          active             true      v26.1.3
4        1          active             true      v26.1.3
5        1          active             true      v26.1.3

Wait for the cluster to stabilize with only green brokers:

kubectl wait --for=condition=Stable redpanda/redpanda -n <namespace> --timeout=600s

Verify the migration

Confirm that only the green brokers remain:

kubectl get pods -n <namespace> -l app.kubernetes.io/name=redpanda

Expected output

NAME               READY   STATUS    RESTARTS   AGE
redpanda-green-0   2/2     Running   0          10m
redpanda-green-1   2/2     Running   0          10m
redpanda-green-2   2/2     Running   0          10m

Verify cluster health:

kubectl exec -n <namespace> redpanda-green-0 -c redpanda -- rpk cluster health

Expected output

CLUSTER HEALTH OVERVIEW
=======================
Healthy:                          true
Unhealthy reasons:                []
Controller ID:                    3
All nodes:                        [3 4 5]
Nodes down:                       []
Leaderless partitions (0):        []
Under-replicated partitions (0):  []

Confirm that all brokers are running the target version:

kubectl exec -n <namespace> redpanda-green-0 -c redpanda -- rpk redpanda admin brokers list

Expected output

NODE-ID  NUM-CORES  MEMBERSHIP-STATUS  IS-ALIVE  BROKER-VERSION
3        1          active             true      v26.1.3
4        1          active             true      v26.1.3
5        1          active             true      v26.1.3

Verify data integrity by checking your topics:

kubectl exec -n <namespace> redpanda-green-0 -c redpanda -- rpk topic list

Confirm that all topics are present and that partition replicas are fully replicated:

kubectl exec -n <namespace> redpanda-green-0 -c redpanda -- rpk topic describe <topic-name>

You can also verify that messages are consumable:

kubectl exec -n <namespace> redpanda-green-0 -c redpanda -- rpk topic consume <topic-name> -n 5

Clean up

Delete the old Kubernetes node pool following your hosting platform’s best practices:

Known limitations

NodePool image tag is required. NodePools do not inherit spec.clusterSpec.image.tag from the Redpanda CR. You must set spec.image.tag on each NodePool explicitly. If not specified, the operator uses its own default image version, which may not be the version you expect.
OnDelete update strategy. NodePool StatefulSets use the OnDelete update strategy. Changing spec.image.tag on an existing NodePool does not trigger an in-place rolling upgrade. Use the blue/green pattern (create a new NodePool) for version upgrades.
Temporary 2x broker count. During migration, the cluster runs double the usual broker count. Ensure your environment has sufficient CPU, memory, and storage resources.