[SOLR-11458] Bugs in MoveReplicaCmd handling of failures

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 7.0, 7.0.1, 7.1, 8.0
• Fix Version/s: 7.2, 8.0
• Assignee: Andrzej Bialecki
• Reporter: Andrzej Bialecki
• Created: 10/Oct/17 10:49
• Updated: 08/Jun/19 15:13
• Resolved: 06/Dec/17 14:34

Description

Spin-off from SOLR-11449:

There's a section of code in moveNormalReplica that ensures that we don't lose a shard leader during move. There's no corresponding protection in moveHdfsReplica, which means that moving a replica that is also a shard leader may potentially lead to data loss (eg. when replicationFactor=1).
Also, there's no rollback strategy when moveHdfsReplica partially fails, unlike in moveNormalReplica where the code simply skips deleting the original replica - it seems that the code should attempt to restore the original replica in this case? When RF=1 and such failure occurs then not restoring the original replica means lost shard.

Issue Links

• is blocked by: SOLR-11661 New HDFS collection reuses unremoved data from a deleted HDFS collection with same name causes inconsistent view of documents