clusterx migrate support data synchronization mode

[yuzegao]

Search before asking

• I had searched in the issues and found no similar issues.

Motivation

After the kvrocks migrate command completes the migration of existing and incremental data, it sets the migrated_slots_ variable and responds to subsequent write requests with the MOVED command, ensuring that subsequent requests are executed only on the migration target node.

In a scenario where we use the migration function to upgrade Kvrocks, we need to observe the data migration to the new kvrocks node for a period of time to ensure that the new node is functioning properly before returning the MOVED command to subsequent write requests. During this observation period, the incremental data synchronization operation is ongoing, meaning that the data is available on both the old and new primary nodes.

Solution

This approach effectively implements phased verification of the new kvrocks program and enables rapid, lossless rollbacks.
To implement the above features, three subcommands need to be added to the clusterx migrate command:
clusterx migrate dtsmode: Data migration mode, continuously synchronizing incremental data
clusterx migrate cancel: Cancels the data migration
clusterx migrate dtscomplete: Completes the migration task, aligning the new primary with the old primary's offset

We will implement this feature. If you have better suggestions, please let us know.

Are you willing to submit a PR?

• I'm willing to submit a PR!

[zhixinwen]

meaning that the data is available on both the old and new primary nodes

Can you clarify how would you send data to both old and new nodes?
I feel I am confused because in the current migration process, there is already a syncWAL phase. How would continuously synchronizing incremental data be different?

[yuzegao]

meaning that the data is available on both the old and new primary nodes

Can you clarify how would you send data to both old and new nodes? I feel I am confused because in the current migration process, there is already a syncWAL phase. How would continuously synchronizing incremental data be different?

As described in the sequence diagram, the original migrate command automatically executes to the KSuccess status.

By adding the DTSMODE and DTSCOMPLETE subcommands, the migration process is split into two phases:

The first phase remains in the KWAL state, enabling continuous writes on the source node and simultaneous synchronization to the target node.

The second phase reaches the KSuccess status, completing data synchronization.