[TAJO-1507] Resource leak when a worker does not response KILLED message

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.11.0
• Component/s: Resource Manager, Worker
• Assignee: Unassigned
• Reporter: Hyunsik Choi
• Created: 02/Apr/15 03:05
• Updated: 04/Aug/15 01:52
• Resolved: 04/Aug/15 01:52

Description

Terminology

• QM - query master
• TM - tajo master

The query kill mechanism is as follows:

• Client sends kill command to TM.
• TM forwards the kill command to QM.
• QM disseminates the kill command to all workers.
• Corresponding workers kill tasks and response KILLED to QM.

BTW, Some workers cannot response KILLED message to QM due to its node failure, temporary network problem, or worker restart. In this case, TM cannot retrieve allocated resources even though the workers become normal.