In the cross-cluster replication validation, we found some cells in source and sink cluster can have the same row key, the same timestamp but different values. The happens when mutations with the same row key are submitted in batch without specifying the timestamp, and the same timestamp in the unit of millisecond is assigned at the time when they are committed to the WAL.
When this happens, if the major compaction hasn’t happened yet and you scan the table, you can find some cells have the same row key, the same timestamps but different values, like the first three rows in the following table.
|Row Key 1||CF0::Column 1||Timestatmp 1||Value 1|
|Row Key 1||CF0::Column 1||Timestatmp 1||Value 2|
|Row Key 1||CF0::Column 1||Timestatmp 1||Value 3|
|Row Key 2||CF0::Column 1||Timestatmp 2||Value 4|
|Row Key 3||CF0::Column 1||Timestatmp 4||Value 5|
The ordering of the first three rows is indeterminate in the presence of the cross-replication, so after compaction, in the master cluster you will see “Row Key 1, CF0::Column1, Timestamp1” having the value 3, but in the slave cluster, you might see the cell having one of the three possible values 1, 2, 3, which results in data inconsistency issue between the master and slave clusters.
Root Cause Analysis:
In HBaseInterClusterReplicationEndpoint.createBatches() of branch-1.3, the WAL entries from the same region could be split into different batches according to replication RPC limit and these batches are shipped by ReplicationSource concurrently, so the batches for the same region could arrive at the sink in the slave clusters then apply to the region synchronously in indeterminate order.
In HBase 3.0.0 and 2.1.0, [~Apache9]&Zheng Hu&Honghua Feng provided Serial Replication
HBASE-20046 which guarantees the order of pushing logs to slave clusters is same as the order of requests from client in the master cluster. It contains mainly two changes:
- Recording the replication "barriers" in ZooKeeper to synchronize the replication across old/failed RS and new RS to provide strict ordering semantics even in the presence of region-move or RS failure.
- Make sure the batches within one region are shipped to the slave clusters in order.
The second part of change is exactly what we need and the minimal change to fix the issue in this JIRA.
To fix the issue in this JIRA, we have two options:
HBASE-20046to branch 1.3. Pros: It also fixes the data inconsistency issue when there is region-move or RS failure and help to reduce the noises in our cross-cluster replication/backup validation which is our ultimate goal. Cons: the change is big and I'm not sure for now whether the change is self-contained or it has other dependencies which need to port to branch 1.3 too; and we need longer time to validate and stabilize.
- Port the minimal change or make the equivalent change as the second part of
HBASE-20046to make sure the batches within one region are shipped to the slave clusters in order."
With limited knowledge about HBase Release Schedule and Process, I prefer option 2 because of cons of option 1, but I'm open to option 1 and other options. Thoughts?