Github user kayousterhout commented on a diff in the pull request:
https://github.com/apache/spark/pull/14079#discussion_r86611855
--- Diff:
core/src/test/scala/org/apache/spark/scheduler/BlacklistTrackerSuite.scala ---
@@ -17,10 +17,299 @@
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 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 (for
thread-safety), 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 { opt =>
+ val actual = f(opt)
+ val exp = expected.contains(opt)
+ assert(actual === exp, raw"""for string "$opt" """)
+ }
+ }
+
+ 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
+ }
-class BlacklistTrackerSuite extends SparkFunSuite {
+ def createTaskSetBlacklist(stageId: Int = 0): TaskSetBlacklist = {
+ new TaskSetBlacklist(conf, stageId, clock)
+ }
+
+ def configureBlacklistAndScheduler(confs: (String, String)*): Unit = {
+ conf = new SparkConf().setAppName("test").setMaster("local")
+ .set(config.BLACKLIST_ENABLED.key, "true")
+ confs.foreach { case (k, v) => conf.set(k, v) }
+ scheduler = mockTaskSchedWithConf(conf)
+
+ clock.setTime(0)
+ blacklist = new BlacklistTracker(conf, clock)
+ }
+
+ test("executors can be blacklisted with only a few failures per stage") {
+ configureBlacklistAndScheduler()
+ // 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 do
not unblacklist the
+ // executor.
+ val failuresTillBlacklisted = conf.get(config.MAX_FAILURES_PER_EXEC)
+ var failuresSoFar = 0
+ (0 until failuresTillBlacklisted * 10).foreach { stage =>
+ val taskSetBlacklist = createTaskSetBlacklist(stage)
+ if (stage % 2 == 0) {
+ // fail every other task
+ taskSetBlacklist.updateBlacklistForFailedTask("hostA", exec = "1",
index = 0)
+ failuresSoFar += 1
+ }
+ blacklist.updateBlacklistForSuccessfulTaskSet(stage, 0,
taskSetBlacklist.execToFailures)
+ assert(failuresSoFar == stage / 2 + 1)
+ if (failuresSoFar < failuresTillBlacklisted) {
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set())
+ } else {
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set("1"))
+ }
+ }
+ }
+
+ // if an executor has many task failures, but the task set ends up
failing, don't count it
+ // against the executor
+ test("executors aren't blacklisted if task sets fail") {
+ configureBlacklistAndScheduler()
+ // for 4 different stages, executor 1 fails a task, and then the
taskSet fails.
+ (0 until 4).foreach { stage =>
+ val taskSetBlacklist = createTaskSetBlacklist(stage)
+ taskSetBlacklist.updateBlacklistForFailedTask("hostA", exec = "1",
index = 0)
+ }
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set())
+ }
+
+ Seq(true, false).foreach { succeedTaskSet =>
+ test(s"stage blacklist updates correctly on stage completion
($succeedTaskSet)") {
+ // within one taskset, an executor fails a few times, so its
blacklisted for the taskset.
+ // but if the taskset fails, we don't blacklist the executor after
the stage.
+ configureBlacklistAndScheduler()
+ val stageId = 1 + (if (succeedTaskSet) 1 else 0)
+ val taskSetBlacklist = createTaskSetBlacklist(stageId)
+ (0 until 4).foreach { index =>
+ taskSetBlacklist.updateBlacklistForFailedTask("hostA", exec = "1",
index = index)
+ }
+ assert(taskSetBlacklist.isExecutorBlacklistedForTaskSet("1"))
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set())
+ if (succeedTaskSet) {
+ // the task set succeeded elsewhere, so we count those failures
against our executor,
+ // and blacklist it across stages
+ blacklist.updateBlacklistForSuccessfulTaskSet(stageId, 0,
taskSetBlacklist.execToFailures)
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set("1"))
+ } else {
+ // the task set failed, so we don't count these failures against
the executor for other
+ // stages
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set())
+ }
+ }
+ }
+
+ test("blacklisted executors and nodes get recovered with time") {
+ configureBlacklistAndScheduler()
+ val taskSetBlacklist0 = createTaskSetBlacklist(stageId = 0)
+ (0 until 4).foreach { partition =>
+ taskSetBlacklist0.updateBlacklistForFailedTask("hostA", exec = "1",
index = partition)
+ }
+ blacklist.updateBlacklistForSuccessfulTaskSet(0, 0,
taskSetBlacklist0.execToFailures)
+ assert(blacklist.nodeBlacklist() === Set())
+ assertEquivalentToSet(blacklist.isNodeBlacklisted(_), Set())
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set("1"))
+
+ val taskSetBlacklist1 = createTaskSetBlacklist(stageId = 1)
+ (0 until 4).foreach { partition =>
+ taskSetBlacklist1.updateBlacklistForFailedTask("hostA", exec = "2",
index = partition)
+ }
+ blacklist.updateBlacklistForSuccessfulTaskSet(0, 0,
taskSetBlacklist1.execToFailures)
+ assert(blacklist.nodeBlacklist() === Set("hostA"))
+ assertEquivalentToSet(blacklist.isNodeBlacklisted(_), Set("hostA"))
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set("1",
"2"))
+
+ clock.advance(blacklist.BLACKLIST_TIMEOUT_MILLIS + 1)
+ blacklist.applyBlacklistTimeout()
+ assert(blacklist.nodeBlacklist() === Set())
+ assertEquivalentToSet(blacklist.isNodeBlacklisted(_), Set())
+ assertEquivalentToSet(blacklist.isExecutorBlacklisted(_), Set())
+
+ // fail one more task, but executor isn't put back into blacklist
since count reset to 0
--- End diff --
since the count of failures on that executor should have been reset to 0
---
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
