[SPARK-39742] Request executor after kill executor, the number of executors is not as expected

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Not A Bug
• Affects Version/s: 3.2.1
• Component/s: Scheduler
• Assignee: Unassigned
• Reporter: Mingliang Zhu
• Created: 11/Jul/22 09:17
• Updated: 26/Oct/22 08:41
• Resolved: 26/Oct/22 08:41

Description

I used the killExecutors and requestExecutors function of SparkContext to dynamically adjust the resources, and found that the requestExecutors after killExecutors could not achieve the expected results.

Add unit tests in StandaloneDynamicAllocationSuite.scala 

test("kill executors first and then request") {
    sc = new SparkContext(appConf
      .set(config.EXECUTOR_CORES, 2)
      .set(config.CORES_MAX, 8))
    val appId = sc.applicationId
    eventually(timeout(10.seconds), interval(10.millis)) {
      val apps = getApplications()
      assert(apps.size === 1)
      assert(apps.head.id === appId)
      assert(apps.head.executors.size === 4) // 8 cores total
      assert(apps.head.getExecutorLimit === Int.MaxValue)
    }
    // sync executors between the Master and the driver, needed because
    // the driver refuses to kill executors it does not know about
    syncExecutors(sc)
    val executors = getExecutorIds(sc)
    assert(executors.size === 4)
    // kill 2 executors
    assert(sc.killExecutors(executors.take(3)))
    val apps = getApplications()
    assert(apps.head.executors.size === 1)
    // add 2 executors
    assert(sc.requestExecutors(3))
    assert(apps.head.executors.size === 4)
  } 

3 did not equal 4
Expected :4
Actual   :3

Issue Links

• links to: [Github] Pull Request #37156 (zml1206)