[YARN-10243] Rack-only localization constraint for MR AM is broken for CapacityScheduler

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 3.2.0
• Component/s: capacity scheduler, capacityscheduler
• Target Version/s: 3.5.0
• Assignee: Bilwa S T
• Reporter: Adam Antal
• Created: 24/Apr/20 11:14
• Updated: 04/Jan/24 08:33

Description

Reproduction: Start a MR sleep job with strict-locality configured for AM (-Dmapreduce.job.am.strict-locality=/rack1 for instance). If CapacityScheduler is used, the job will hang (stuck in SCHEDULED state).

Root cause: if there are no other resources requested (like node locality or other constraint), the scheduling opportunities counter will not be incremented and the following piece of code always returns false (so we always skip this constraint) resulting in an infinite loop:

    // If we are here, we do need containers on this rack for RACK_LOCAL req
    if (type == NodeType.RACK_LOCAL) {
      // 'Delay' rack-local just a little bit...
      long missedOpportunities =
          application.getSchedulingOpportunities(schedulerKey);
      return getActualNodeLocalityDelay() < missedOpportunities;
    }

Workaround: set yarn.scheduler.capacity.node-locality-delay to zero to enforce this rule to be processed immediately.