[HDFS-11889] Datanodes should skip reading replicas from cache during Rollback

• Type: Bug
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Component/s: datanode, rolling upgrades
• Assignee: Vinayakumar B
• Reporter: Vinayakumar B
• Created: 26/May/17 08:57
• Updated: 28/May/17 14:55

Description

During rolling upgrade rollback, Datanodes should skip reading from replica cache file.
Cache file will be written if
1. Datanodes shutdown gracefully by dfsadmin command.
2. For tests, Datanode.shutdown() is called.

Cache file will not have entries of blocks restored from trash. So these blocks will not be reported back to Namenode until another directory scan happens.

So avoid reading from cache even if exists if ROLLBACK option is used.