[YARN-11014] YARN incorrectly validates maximum capacity resources on the validation API

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.4.0, 3.2.4, 3.3.3
• Fix Version/s: 3.4.0, 3.2.4, 3.3.3
• Component/s: capacity scheduler
• Labels: pull-request-available
• Target Version/s: 3.3.3
• Hadoop Flags: Reviewed
• Assignee: Benjamin Teke
• Reporter: Benjamin Teke
• Created: 23/Nov/21 15:43
• Updated: 11/Feb/24 23:42
• Resolved: 18/Apr/22 14:13

Description

The YARN validation API (http://<host>:8088/ws/v1/cluster/scheduler-conf/validate) allows a configuration in which a child queue has a greater maximum capacity value than its parent (the same update fails when it is attempted for real) using the DominantResourceCalculator.
For example, the following passes validation when the root queue's maximum capacity is less than the 100000000 it is attempting to set here:

{
    "update-queue": [
        {
            "queue-name": "root.test",
            "params": {
                "entry": [
                    {
                        "key": "maximum-capacity",
                        "value": "[memory=100000000,vcores=3]"
                    },
                    {
                        "key": "capacity",
                        "value": "[memory=3020,vcores=1]"
                    }
                ]
            }
        }
    ]
} 

The reason for this is the newly created CapacityScheduler instance doesn't have the active nodes of the cluster in the nodeTracker object, hence the clusterResources will be Resources.none() during the init/reinit phase of this new instance.
If the dominantRC has invalid denominator (clusterResource being 0) it simply compares the two values resource-by-resource:

  @Override
  public int compare(Resource clusterResource, Resource lhs, Resource rhs,
      boolean singleType) {
    if (lhs.equals(rhs)) {
      return 0;
    }

    if (isAllInvalidDivisor(clusterResource)) {
      return this.compare(lhs, rhs);
    }
...

  /**
   * Compare two resources - if the value for every resource type for the lhs
   * is greater than that of the rhs, return 1. If the value for every resource
   * type in the lhs is less than the rhs, return -1. Otherwise, return 0
   *
   * @param lhs resource to be compared
   * @param rhs resource to be compared
   * @return 0, 1, or -1
   */
   private int compare(Resource lhs, Resource rhs) {

Issue Links

• is depended upon by: HADOOP-18198 Release Hadoop 3.3.3: hadoop-3.3.2 with some fixes
• requires: YARN-10503 Support queue capacity in terms of absolute resources with custom resourceType.
• links to: GitHub Pull Request #3715
• links to: GitHub Pull Request #3772
• links to: GitHub Pull Request #3773