Remove a node

To reduce the size of a cluster, you can remove nodes. This process removes the node without replacing it with a new node. To replace a node, see Replace a running node or Replace a dead node.

Use nodetool status to determine if the node is up (UN) or down (DN).

For example:

Datacenter: datacenter1
================================
Status=Up/Down
|/ State=Normal/Leaving/Joining/Moving
--  Address        Load       Tokens       Owns    Host ID                               Rack
UN  10.200.177.92  265.04 KiB  1            ?       980cab6a-2e5d-44c6-b897-0733dde580ac  rack1
DN  10.200.177.94  426.21 KiB  1            ?       7ecbbc0c-627d-403e-b8cc-a2daa93d9ad3  rack1

If the node is up, run nodetool decommission.

This command reassigns the node’s token ranges and replicates the node’s data appropriately before the node is removed from the cluster. To avoid excessive data streaming, decommission only one node at a time. You can use nodetool netstats to monitor the progress of decommissioning.

If nodetool decommission succeeds, stop HCD and shut down the node to finalize the removal.

If nodetool decommission fails, proceed to the next step.

If the node is down or nodetool decommission fails, remove the node according to your cluster architecture:

Virtual nodes (vnodes): Run nodetool removenode. Vnodes are automatically redistributed across the remaining nodes.

Single-token architecture: Manually redistribute tokens across the remaining nodes by editing initial_token in cassandra.yaml, and then run nodetool removenode.

In single-token architecture clusters (not using vnodes), you must manually rebalance the cluster before removing a node. If you don’t redistribute the tokens, the next node in the ring will take all tokens from the removed node, resulting in imbalanced data distribution and potential hot spots.

If nodetool removenode fails, run nodetool assassinate.

Remove a node

Was this helpful?

Give Feedback