A new node under normal operation joins a ring in the UJ (Up Joining) state.
Gossip between the nodes report the new node’s addition and the nodes wait for the new node to change to the UN (Up Normal) state.
Data is re-distributed to include the new node and when it has its share of the data, the new node’s status will change to UN (Up Normal) state.
Should this fail to happen, the node may join the cluster directly into the UN state without receiving any data.
Use the procedure that follows to correct situations where a node joins a cluster directly in the UN state.
Note
This procedure only covers the situation in which a new node joins the cluster. Nodes replacing existing nodes following the Replace A Dead Node procedure are expected to be in UN state from the start.
There are two possible causes for a new node not acquiring data from its peers:
The new node’s configuration file (scylla.yml) lists only itself as the seed node and doesn’t use the same seed address as the other nodes in the cluster do.
The node was added with the auto_bootstrap flag set to False
Both scenarios constitute user errors and are the incorrect way of adding new nodes to the cluster. See Adding New Node for more details. This article explains how to restore a cluster safely in such a case.
Warning
You mustn’t run nodetool cleanup in any other nodes in the cluster until this is fixed, as it will delete data that is not yet available from the new node.
Steps to fix:
For every node that joined the cluster directly in UN state, run nodetool decommission serially. Do not decommission more than one node at the same time.
Make sure to clean the data from the decommissioned node
sudo rm -rf /var/lib/scylla/data
sudo find /var/lib/scylla/commitlog -type f -delete
sudo find /var/lib/scylla/hints -type f -delete
sudo find /var/lib/scylla/view_hints -type f -delete
Actual directories might be different than above, check your scylla.yaml to validate.
Once all new nodes are decommissioned, add them again with the correct procedure. Do not add nodes as seeds. Do not add nodes with the auto_bootstrap flag set to False.
Other important considerations:
Avoid running nodetool removenode for the new nodes. Running nodetool removenode instead of nodetool decommission will cause the new writes sent to the new nodes to be lost. In all cases, you should always prefer to decommission over removenode. Use removenode only in cases where there is no other way to restart the node.
While running a repair in the cluster will bring some data back to the new node and restore its data, it will take much longer than the procedure suggested here. It is also important to keep in mind that one copy is lost per node inserted directly into the UN state. In the extreme case in which N nodes were added where N is your replication factor, some data may have been permanently lost and repairs will not be able to restore it. Therefore, repair is not recommended to fix this issue, except when the user is aware and their consistency model can tolerate missing data points.