[AMBARI-20682] Wait For DataNodes To Shutdown During a Rolling Upgrade

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 2.4.2
• Fix Version/s: 2.5.1
• Component/s: ambari-server
• Assignee: Dmitry Lysnichenko
• Reporter: Dmitry Lysnichenko
• Created: 05/Apr/17 12:25
• Updated: 16/Feb/18 13:14
• Resolved: 13/Apr/17 14:38

Description

During a rolling upgrade (especially on a large, heavily used cluster), the DataNodes do not shutdown immediately. However, they do de-register from the NameNode which tricks Ambari into thinking that they are down.

Since the rolling upgrade uses a RESTART command, we attempt to start the DataNode back up before the daemon has shutdown:

2017-03-14 05:00:25,602 - call['/usr/hdp/current/hadoop-hdfs-datanode/bin/hdfs dfsadmin -fs hdfs://c1ha -shutdownDatanode 0.0.0.0:8010 upgrade'] {'user': 'hdfs'}
2017-03-14 05:00:28,438 - call returned (0, 'Submitted a shutdown request to datanode 0.0.0.0:8010')
2017-03-14 05:00:28,438 - Execute['/usr/hdp/current/hadoop-hdfs-datanode/bin/hdfs dfsadmin -fs hdfs://c1ha -D ipc.client.connect.max.retries=5 -D ipc.client.connect.retry.interval=1000 -getDatanodeInfo 0.0.0.0:8010'] {'tries': 1, 'user': 'hdfs'}
2017-03-14 05:00:35,976 - DataNode has successfully shutdown for upgrade.

Even though ~ 6 seconds have passed, the daemon is still running as it drains. Therefore, we attempt to start it which causes a NOOP.

Instead, we should also monitor for the PID.

Issue Links

• links to: review_request_