-
Notifications
You must be signed in to change notification settings - Fork 59
Open
Description
In the tryUpdateMigrationStatus
logic, after a successful migration, the controller first updates the topology in the kvrocks cluster via SetSlot
, then updates the etcd data using clusterStore.UpdateCluster
. However, there appears to be a potential vulnerability here:
- If
SetSlot
succeeds but a network failure preventsUpdateCluster
from completing (or worse, ifSetSlot
partially succeeds with some nodes updated and others failing), this could leave the kvrocks cluster version higher than the cluster version stored in etcd. - Even if another controller takes over from the failed one, it would be unable to execute new operations because subsequent "versions" issued by the controller would no longer exceed the version already in kvrocks.
To address this, I suggest reversing the order: update etcd first, then propagate the new topology to kvrocks.
As a side note, if this approach is adopted, the clusterx set slot
command in kvrocks might become redundant, as we could fully update the post-migration topology using setnodes
instead.
Metadata
Metadata
Assignees
Labels
No labels