Github user kayousterhout commented on a diff in the pull request: https://github.com/apache/spark/pull/14079#discussion_r92079858 --- Diff: core/src/test/scala/org/apache/spark/scheduler/BlacklistTrackerSuite.scala --- @@ -17,10 +17,322 @@ package org.apache.spark.scheduler -import org.apache.spark.{SparkConf, SparkFunSuite} +import org.mockito.Mockito.when +import org.scalatest.BeforeAndAfterEach +import org.scalatest.mock.MockitoSugar + +import org.apache.spark._ import org.apache.spark.internal.config +import org.apache.spark.util.ManualClock + +class BlacklistTrackerSuite extends SparkFunSuite with BeforeAndAfterEach with MockitoSugar + with LocalSparkContext { + + private val clock = new ManualClock(0) + + private var blacklist: BlacklistTracker = _ + private var scheduler: TaskSchedulerImpl = _ + private var conf: SparkConf = _ + + override def beforeEach(): Unit = { + conf = new SparkConf().setAppName("test").setMaster("local") + .set(config.BLACKLIST_ENABLED.key, "true") + scheduler = mockTaskSchedWithConf(conf) + + clock.setTime(0) + blacklist = new BlacklistTracker(conf, clock) + } + + override def afterEach(): Unit = { + if (blacklist != null) { + blacklist = null + } + if (scheduler != null) { + scheduler.stop() + scheduler = null + } + super.afterEach() + } + + val allExecutorAndHostIds = (('A' to 'Z').map("host" + _) ++ (1 to 100).map{_.toString}).toSet + + /** + * Its easier to write our tests as if we could directly look at the sets of nodes & executors in + * the blacklist. However the api doesn't expose a set, so this is a simple way to test + * something similar, since we know the universe of values that might appear in these sets. + */ + def assertEquivalentToSet(f: String => Boolean, expected: Set[String]): Unit = { + allExecutorAndHostIds.foreach { id => + val actual = f(id) + val exp = expected.contains(id) + assert(actual === exp, raw"""for string "$id" """) + } + } + + def mockTaskSchedWithConf(conf: SparkConf): TaskSchedulerImpl = { + sc = new SparkContext(conf) + val scheduler = mock[TaskSchedulerImpl] + when(scheduler.sc).thenReturn(sc) + when(scheduler.mapOutputTracker).thenReturn(SparkEnv.get.mapOutputTracker) + scheduler + } + + def createTaskSetBlacklist(stageId: Int = 0): TaskSetBlacklist = { + new TaskSetBlacklist(conf, stageId, clock) + } + + test("executors can be blacklisted with only a few failures per stage") { + // For 4 different stages, executor 1 fails a task, then executor 2 succeeds the task, + // and then the task set is done. Not enough failures to blacklist the executor *within* + // any particular taskset, but we still blacklist the executor overall eventually. + // Also, we intentionally have a mix of task successes and failures -- there are even some + // successes after the executor is blacklisted. The idea here is those tasks get scheduled + // before the executor is blacklisted. We might get successes after blacklisting (because the + // executor might be flaky but not totally broken). But successes should not unblacklist the + // executor. + val failuresUntilBlacklisted = conf.get(config.MAX_FAILURES_PER_EXEC) + var failuresSoFar = 0 + (0 until failuresUntilBlacklisted * 10).foreach { stageId => + val taskSetBlacklist = createTaskSetBlacklist(stageId) + if (stageId % 2 == 0) { + // fail every other task --- End diff -- fail a task in every other stage?
--- If your project is set up for it, you can reply to this email and have your reply appear on GitHub as well. If your project does not have this feature enabled and wishes so, or if the feature is enabled but not working, please contact infrastructure at infrastruct...@apache.org or file a JIRA ticket with INFRA. --- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org