[GOBBLIN-780] Handle scenarios that cause the YarnAutoScalingManager to be stuck

• Type: Task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.15.0
• Assignee: Unassigned
• Reporter: Hung Tran
• Created: 23/May/19 16:17
• Updated: 28/May/19 16:56
• Resolved: 28/May/19 16:56

Description

Issue 1: The YarnAutoScalingRunnable is run in a fixed schedule by a ScheduledExecutorService in YarnAutoScalingManager. If the runnable encounters an exception the the executor service will stop scheduling it. Catch all exceptions in the runnable, log, and do not re-raise.

Issue 2: The auto scaler may reduce the container count to 0. Helix will not schedule any flows if there are no participants connected. This results in the auto scaler keeping the container count at 0 and no progress is made. Fix this by not allowing the container count to be reduced below 1.

 

 

Issue Links

• links to: GitHub Pull Request #2644