Robert Kanter created YARN-5465:
-----------------------------------
Summary: Server-Side NM Graceful Decommissioning subsequent call
behavior
Key: YARN-5465
URL: https://issues.apache.org/jira/browse/YARN-5465
Project: Hadoop YARN
Issue Type: Sub-task
Reporter: Robert Kanter
The Server-Side NM Graceful Decommissioning feature added by YARN-4676 has the
following behavior when subsequent calls are made:
# Start a long-running job that has containers running on nodeA
# Add nodeA to the exclude file
# Run {{-refreshNodes -g 120 -server}} (2min) to begin gracefully
decommissioning nodeA
# Wait 30 seconds
# Add nodeB to the exclude file
# Run {{-refreshNodes -g 30 -server}} (30sec)
# After 30 seconds, both nodeA and nodeB shut down
In a nutshell, issuing a subsequent call to gracefully decommission nodes
updates the timeout for any currently decommissioning nodes. This makes it
impossible to gracefully decommission different sets of nodes with different
timeouts. Though it does let you easily update the timeout of currently
decommissioning nodes.
Another behavior we could do is this:
# {color:grey}Start a long-running job that has containers running on nodeA
# {color:grey}Add nodeA to the exclude file{color}
# {color:grey}Run {{-refreshNodes -g 120 -server}} (2min) to begin gracefully
decommissioning nodeA{color}
# {color:grey}Wait 30 seconds{color}
# {color:grey}Add nodeB to the exclude file{color}
# {color:grey}Run {{-refreshNodes -g 30 -server}} (30sec){color}
# After 30 seconds, nodeB shuts down
# After 60 more seconds, nodeA shuts down
This keeps the nodes affected by each call to gracefully decommission nodes
independent. You can now have different sets of decommissioning nodes with
different timeouts. However, to update the timeout of a currently
decommissioning node, you'd have to first recommission it, and then
decommission it again.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
