[MAPREDUCE-1485] CapacityScheduler should have prevent a single job taking over large parts of a cluster

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Component/s: capacity-sched
• Assignee: Arun Murthy
• Reporter: Arun Murthy
• Created: 11/Feb/10 19:52
• Updated: 29/Jul/14 23:31
• Resolved: 29/Jul/14 23:31

Description

The proposal is to have a per-queue limit on the number of concurrent tasks a job can run on a cluster.

We've seen cases where a single, large, job took over a majority of the cluster - worse, it meant that any bug in it caused issues for both the NameNode and the JobTracker.