[HDFS-7442] Optimization for decommission-in-progress check

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 2.6.0
• Component/s: namenode
• Assignee: Unassigned
• Reporter: Ming Ma
• Created: 25/Nov/14 07:15
• Updated: 01/May/15 19:52
• Resolved: 01/May/15 19:52

Description

1. {{isReplicationInProgress }} currently rescan all blocks of a given node each time the method is called; it becomes less efficient as more of its blocks become fully replicated. Each scan takes FS lock.

2. As discussed in HDFS-7374, if the node becomes dead during decommission, it is useful if the dead node can be marked as decommissioned after all its blocks are fully replicated. Currently there is no way to check the blocks of dead decomm-in-progress nodes, given the dead node is removed from blockmap.

There are mitigations for these limitations. Set dfs.namenode.decommission.nodes.per.interval to small value for reduce the duration of lock. HDFS-7409 uses global FS state to tell if a dead node's blocks are fully replicated.

To address these scenarios, it will be useful to track the decommon-in-progress blocks separately.

Issue Links

• duplicates: HDFS-7411 Refactor and improve decommissioning logic into DecommissionManager