Was this page helpful?
Caution
You're viewing documentation for an unstable version of ScyllaDB Enterprise. Switch to the latest stable version.
Note
This page only applies to clusters where consistent topology updates are not enabled. The page does NOT apply if you:
Created a cluster with ScyllaDB 2024.1 (consistent topology updates are automatically enabled).
Upgraded from ScyllaDB 2024.1 and manually enabled consistent topology updates.
A failure may happen in the middle of a cluster membership change (that is bootstrap, decommission, removenode, or replace), such as loss of power. If that happens, you should ensure that the cluster is brought back to a consistent state as soon as possible. Further membership changes might be impossible until you do so.
For example, a node that crashed in the middle of decommission might leave the cluster in a state where it considers the node to still be a member, but the node itself will refuse to restart and communicate with the cluster. This particular case is very unlikely - it requires a specifically timed crash to happen, after the data streaming phase of decommission finishes but before the node commits that it left. But if it happens, you won’t be able to bootstrap other nodes (they will try to contact the partially-decommissioned node and fail) until you remove the remains of the node that crashed.
If a failure happens when trying to bootstrap a new node to the cluster, you can try bootstrapping the node again by restarting it.
If the failure persists or you decided that you don’t want to bootstrap the node anymore, follow the instructions in the cleaning up after a failed membership change section to remove the remains of the bootstrapping node. You can then clear the node’s data directories and attempt to bootstrap it again.
There are two cases.
Most likely the failure happened during the data repair/streaming phase - before the node tried to leave the token ring. Look for a log message containing “leaving token ring” in the logs of the node that you tried to decommission. For example:
INFO 2023-03-14 13:08:38,323 [shard 0] storage_service - decommission[5b2e752e-964d-4f36-871f-254491f4e8cc]: leaving token ring
If the message is not present, the failure happened before the node tried to leave the token ring. In that case you can simply restart the node and attempt to decommission it again.
If the message is present, the node attempted to leave the token ring, but it might have left the cluster only partially before the failure. Do not try to restart the node. Instead, you must make sure that the node is dead and remove any leftovers using the removenode operation. See cleaning up after a failed membership change. Trying to restart the node after such failure results in unpredictable behavior - it may restart normally, it may refuse to restart, or it may even try to rebootstrap.
If you don’t have access to the node’s logs anymore, assume the second case (the node might have attempted to leave the token ring), do not try to restart the node, instead follow the cleaning up after a failed membership change section.
Simply retry the removenode operation.
If you somehow lost the host ID of the node that you tried to remove, follow the instructions in cleaning up after a failed membership change.
Replace is a special case of bootstrap, but the bootstrapping node tries to take the place of another dead node. You can retry a failed replace operation by restarting the replacing node.
If the failure persists or you decided that you don’t want to perform the replace anymore, follow the instructions in cleaning up after a failed membership change section to remove the remains of the replacing node. You can then clear the node’s data directories and attempt to replace again. Alternatively, you can remove the dead node which you initially tried to replace using removenode, and perform a regular bootstrap.
After a failed membership change, the cluster may contain remains of a node that tried to leave or join - other nodes may consider the node a member, possibly in a transitioning state. It is important to remove any such “ghost” members. Their presence may reduce the cluster’s availability, performance, or prevent further membership changes.
You need to determine the host IDs of any potential ghost members, then remove them using the removenode operation. Note that after a failed replace, there may be two different host IDs that you’ll want to find and run removenode
on: the new replacing node and the old node that you tried to replace. (Or you can remove the new node only, then try to replace the old node again.)
After a failed bootstrap, you need to determine the host ID of the node that tried to bootstrap, if it managed to generate a host ID (it might not have chosen the host ID yet if it failed very early in the procedure, in which case there’s nothing to remove). Look for a message containing system_keyspace - Setting local host id to
in the node’s logs, which will contain the node’s host ID. For example: system_keyspace - Setting local host id to f180b78b-6094-434d-8432-7327f4d4b38d
. If you don’t have access to the node’s logs, read the generic method below.
After a failed decommission, you need to determine the host ID of the node that tried to decommission. You can search the node’s logs as in the failed bootstrap case (see above), or you can use the generic method below.
After a failed removenode, you need to determine the host ID of the node that you tried to remove. You should already have it, since executing a removenode requires the host ID in the first place. But if you lost it somehow, read the generic method below.
After a failed replace, you need to determine the host ID of the replacing node. Search the node’s logs as in the failed bootstrap case (see above), or you can use the generic method below. You may also want to determine the host ID of the replaced node - either to attempt replacing it again after removing the remains of the previous replacing node, or to remove it using nodetool removenode. You should already have the host ID of the replaced node if you used the replace_node_first_boot
option to perform the replace.
If you cannot determine the ghost members’ host ID using the suggestions above, use the method described below.
Make sure there are no ongoing membership changes.
Execute the following CQL query on one of your nodes to retrieve the Raft group 0 ID:
select value from system.scylla_local where key = 'raft_group0_id'
For example:
cqlsh> select value from system.scylla_local where key = 'raft_group0_id';
value
--------------------------------------
607fef80-c276-11ed-a6f6-3075f294cc65
Use the obtained Raft group 0 ID to query the set of all cluster members’ host IDs (which includes the ghost members), by executing the following query:
select server_id from system.raft_state where group_id = <group0_id>
replace <group0_id>
with the group 0 ID that you obtained. For example:
cqlsh> select server_id from system.raft_state where group_id = 607fef80-c276-11ed-a6f6-3075f294cc65;
server_id
--------------------------------------
26a9badc-6e96-4b86-a8df-5173e5ab47fe
7991e7f5-692e-45a0-8ae5-438be5bc7c4f
aff11c6d-fbe7-4395-b7ca-3912d7dba2c6
Execute the following CQL query to obtain the host IDs of all token ring members:
select host_id, up from system.cluster_status;
For example:
cqlsh> select peer, host_id, up from system.cluster_status;
peer | host_id | up
-----------+--------------------------------------+-------
127.0.0.3 | null | False
127.0.0.1 | 26a9badc-6e96-4b86-a8df-5173e5ab47fe | True
127.0.0.2 | 7991e7f5-692e-45a0-8ae5-438be5bc7c4f | True
The output of this query is similar to the output of nodetool status
.
We included the up
column to see which nodes are down and the peer
column to see their IP addresses.
In this example, one of the nodes tried to decommission and crashed as soon as it left the token ring but before it left the Raft group. Its entry will show up in system.cluster_status
queries with host_id = null
, like above, until the cluster is restarted.
A host ID belongs to a ghost member if:
It appears in the system.raft_state
query but not in the system.cluster_status
query,
Or it appears in the system.cluster_status
query but does not correspond to any remaining node in your cluster.
In our example, the ghost member’s host ID was aff11c6d-fbe7-4395-b7ca-3912d7dba2c6
because it appeared in the system.raft_state
query but not in the system.cluster_status
query.
If you’re unsure whether a given row in the system.cluster_status
query corresponds to a node in your cluster, you can connect to each node in the cluster and execute select host_id from system.local
(or search the node’s logs) to obtain that node’s host ID, collecting the host IDs of all nodes in your cluster. Then check if each host ID from the system.cluster_status
query appears in your collected set; if not, it’s a ghost member.
A good rule of thumb is to look at the members marked as down (up = False
in system.cluster_status
) - ghost members are eventually marked as down by the remaining members of the cluster. But remember that a real member might also be marked as down if it was shutdown or partitioned away from the rest of the cluster. If in doubt, connect to each node and collect their host IDs, as described in the previous paragraph.
In some cases, even after a failed topology change, there may be no ghost members left - for example, if a bootstrapping node crashed very early in the procedure or a decommissioning node crashed after it committed the membership change but before it finalized its own shutdown steps.
If any ghost members are present, proceed to the next step.
Given the host IDs of ghost members, you can remove them using removenode
; follow the documentation for removenode operation.
If you’re executing removenode
too quickly after a failed membership change, an error similar to the following might pop up:
nodetool: ScyllaDB API server HTTP POST to URL '/storage_service/remove_node' failed: seastar::rpc::remote_verb_error (node_ops_cmd_check: Node 127.0.0.2 rejected node_ops_cmd=removenode_abort from node=127.0.0.1 with ops_uuid=0ba0a5ab-efbd-4801-a31c-034b5f55487c, pending_node_ops={b47523f2-de6a-4c38-8490-39127dba6b6a}, pending node ops is in progress)
In that case simply wait for 2 minutes before trying removenode
again.
If removenode
returns an error like:
nodetool: ScyllaDB API server HTTP POST to URL '/storage_service/remove_node' failed: std::runtime_error (removenode[12e7e05b-d1ae-4978-b6a6-de0066aa80d8]: Host ID 42405b3b-487e-4759-8590-ddb9bdcebdc5 not found in the cluster)
and you’re sure that you’re providing the correct Host ID, it means that the member was already removed and you don’t have to clean up after it.
Was this page helpful?
On this page