[YARN-11682] Legacy auto created queue in absolute mode has zero capacity after creation during app recovery

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: capacity scheduler
• Assignee: Susheel Gupta
• Reporter: Brian Goerlitz
• Created: 10/Apr/24 02:59
• Updated: 2 days ago 05:50

Description

During recovery of a running app in a legacy auto created queue configured in absolute mode the configured min resources will be set to zero as NodeManagers have not registered yet (clusterResource is zero).

GuaranteedOrZeroCapacityOverTimePolicy.getInitialLeafQueueConfiguration(AbstractAutoCreatedLeafQueue leafQueue)
...
   if (availableCapacity >= leafQueueTemplateCapacities
            .getAbsoluteCapacity(nodeLabel)) {
          updateCapacityFromTemplate(capacities, nodeLabel);
          activate(leafQueue, nodeLabel);
        } else{
          updateToZeroCapacity(capacities, nodeLabel, leafQueue);
        }

GuaranteedOrZeroCapacityOverTimePolicy.updateToZeroCapacity(QueueCapacities capacities, String nodeLabel, AbstractLeafQueue leafQueue)

private void updateToZeroCapacity(QueueCapacities capacities,
      String nodeLabel, AbstractLeafQueue leafQueue) {
    capacities.setCapacity(nodeLabel, 0.0f);
    capacities.setMaximumCapacity(nodeLabel,
        leafQueueTemplateCapacities.getMaximumCapacity(nodeLabel));
    leafQueue.getQueueResourceQuotas().
        setConfiguredMinResource(nodeLabel, Resource.newInstance(0, 0));
  }

When a NodeManager is registered, AbstractCSQueue.updateEffectiveResources(Resource clusterResource) is called, but specifically absolute mode queues are updated using the configured min resource, which will now be zero.

AbstractCSQueue.updateEffectiveResources(Resource clusterResource)
...
if (getCapacityConfigType().equals(
          CapacityConfigType.ABSOLUTE_RESOURCE)) {
        newEffectiveMinResource = createNormalizedMinResource(
            usageTracker.getQueueResourceQuotas().getConfiguredMinResource(label),
            ((AbstractParentQueue) parent).getEffectiveMinRatio(label));
...
   usageTracker.getQueueResourceQuotas().setEffectiveMinResource(label,
          newEffectiveMinResource);

Reinitializing the queue via a config change will correctly recalculate the capacity based on current clusterResource.