[HBASE-14004] [Replication] Inconsistency between Memstore and WAL may result in data in remote cluster that is not in the origin

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 2.0.0-alpha-4, 2.0.0
• Component/s: regionserver, Replication
• Labels: replication wal
• Hadoop Flags: Reviewed

Release Note

Now when replicating a wal file which is still opened for write, we will get its committed length from the WAL instance in the same RS to prevent replicating uncommit WALEdit.

This is very important if you use AsyncFSWAL, as we use fan-out in AsyncFSWAL. The data written to DN will be visible immediately as all DNs think it is the end of a pipeline, although the client has not received an ack, and also NN may truncate the file if the client crashes at the same time.

• Assignee: Duo Zhang
• Reporter: He Liangliang
• Created: 01/Jul/15 12:32
• Updated: 01/Nov/23 22:46
• Resolved: 15/Sep/17 12:34

Description

Looks like the current write path can cause inconsistency between memstore/hfile and WAL which cause the slave cluster has more data than the master cluster.

The simplified write path looks like:
1. insert record into Memstore
2. write record to WAL
3. sync WAL
4. rollback Memstore if 3 fails

It's possible that the HDFS sync RPC call fails, but the data is already (may partially) transported to the DNs which finally get persisted. As a result, the handler will rollback the Memstore and the later flushed HFile will also skip this record.

==================================

This is a long lived issue. The above problem is solved by write path reorder, as now we will sync wal first before modifying memstore. But the problem may still exists as replication thread may read the new data before we return from hflush. See this document for more details:

https://docs.google.com/document/d/11AyWtGhItQs6vsLRIx32PwTxmBY3libXwGXI25obVEY/edit#

So we need to keep a sync length in WAL and tell replication wal reader this is limit when you read this wal file.

Issue Links

• breaks: HBASE-21503 Replication normal source can get stuck due potential race conditions between source wal reader and wal provider initialization threads.
• breaks: HBASE-18845 TestReplicationSmallTests fails after HBASE-14004
• is related to: HBASE-24625 AsyncFSWAL.getLogFileSizeIfBeingWritten does not return the expected synced file length.
• is related to: HBASE-28184 Tailing the WAL is very slow if there are multiple peers.
• relates to: HBASE-5954 Allow proper fsync support for HBase
• relates to: HBASE-14790 Implement a new DFSOutputStream for logging WAL only
• links to: fsync behind and fsync on close (and HDFS-744)
• links to: Review Board
• links to: The new design for HBase writing logic to prevent data loss and inconsistency(HBASE-14004)