Skip to content
Snippets Groups Projects
Commit 981fbafa authored by Marcelo Vanzin's avatar Marcelo Vanzin Committed by Sean Owen
Browse files

[SPARK-6325] [core,yarn] Do not change target executor count when killing executors.

The dynamic execution code has two ways to reduce the number of executors: one
where it reduces the total number of executors it wants, by asking for an absolute
number of executors that is lower than the previous one. The second is by
explicitly killing idle executors.

YarnAllocator was mixing those up and lowering the target number of executors
when a kill was issued. Instead, trust the frontend knows what it's doing, and kill
executors without messing with other accounting. That means that if the frontend
kills an executor without lowering the target, it will get a new executor shortly.

The one situation where both actions (lower the target and kill executor) need to
happen together is when user code explicitly calls `SparkContext.killExecutors`.
In that case, issue two calls to the backend to achieve the goal.

I also did some minor cleanup in related code:
- avoid sending a request for executors when target is unchanged, to avoid log
  spam in the AM
- avoid printing misleading log messages in the AM when there are no requests
  to cancel
- fix a slow memory leak plus misleading error message on the driver caused by
  failing to completely unregister the executor.

Author: Marcelo Vanzin <vanzin@cloudera.com>

Closes #5018 from vanzin/SPARK-6325 and squashes the following commits:

2e782a3 [Marcelo Vanzin] Avoid redundant logging on the AM side.
a3567cd [Marcelo Vanzin] Add parentheses.
a363926 [Marcelo Vanzin] Update logic.
a158101 [Marcelo Vanzin] [SPARK-6325] [core,yarn] Disallow reducing executor count past running count.
parent 9d112a95
No related branches found
No related tags found
No related merge requests found
......@@ -211,6 +211,7 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val actorSyste
// This must be synchronized because variables mutated
// in this block are read when requesting executors
CoarseGrainedSchedulerBackend.this.synchronized {
addressToExecutorId -= executorInfo.executorAddress
executorDataMap -= executorId
executorsPendingToRemove -= executorId
}
......@@ -371,6 +372,12 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, val actorSyste
logWarning(s"Executor to kill $id does not exist!")
}
}
// Killing executors means effectively that we want less executors than before, so also update
// the target number of executors to avoid having the backend allocate new ones.
val newTotal = (numExistingExecutors + numPendingExecutors - executorsPendingToRemove.size
- filteredExecutorIds.size)
doRequestTotalExecutors(newTotal)
executorsPendingToRemove ++= filteredExecutorIds
doKillExecutors(filteredExecutorIds)
}
......
......@@ -534,7 +534,6 @@ private[spark] class ApplicationMaster(
driver ! x
case RequestExecutors(requestedTotal) =>
logInfo(s"Driver requested a total number of $requestedTotal executor(s).")
Option(allocator) match {
case Some(a) => a.requestTotalExecutors(requestedTotal)
case None => logWarning("Container allocator is not ready to request executors yet.")
......
......@@ -86,7 +86,8 @@ private[yarn] class YarnAllocator(
@volatile private var targetNumExecutors = args.numExecutors
// Keep track of which container is running which executor to remove the executors later
private val executorIdToContainer = new HashMap[String, Container]
// Visible for testing.
private[yarn] val executorIdToContainer = new HashMap[String, Container]
// Executor memory in MB.
protected val executorMemory = args.executorMemory
......@@ -137,7 +138,10 @@ private[yarn] class YarnAllocator(
* be killed.
*/
def requestTotalExecutors(requestedTotal: Int): Unit = synchronized {
targetNumExecutors = requestedTotal
if (requestedTotal != targetNumExecutors) {
logInfo(s"Driver requested a total number of $requestedTotal executor(s).")
targetNumExecutors = requestedTotal
}
}
/**
......@@ -148,8 +152,6 @@ private[yarn] class YarnAllocator(
val container = executorIdToContainer.remove(executorId).get
internalReleaseContainer(container)
numExecutorsRunning -= 1
targetNumExecutors -= 1
assert(targetNumExecutors >= 0, "Allocator killed more executors than are allocated!")
} else {
logWarning(s"Attempted to kill unknown executor $executorId!")
}
......@@ -351,7 +353,8 @@ private[yarn] class YarnAllocator(
}
}
private def processCompletedContainers(completedContainers: Seq[ContainerStatus]): Unit = {
// Visible for testing.
private[yarn] def processCompletedContainers(completedContainers: Seq[ContainerStatus]): Unit = {
for (completedContainer <- completedContainers) {
val containerId = completedContainer.getContainerId
......
......@@ -206,6 +206,28 @@ class YarnAllocatorSuite extends FunSuite with Matchers with BeforeAndAfterEach
handler.getNumExecutorsRunning should be (2)
}
test("kill executors") {
val handler = createAllocator(4)
handler.updateResourceRequests()
handler.getNumExecutorsRunning should be (0)
handler.getNumPendingAllocate should be (4)
val container1 = createContainer("host1")
val container2 = createContainer("host2")
handler.handleAllocatedContainers(Array(container1, container2))
handler.requestTotalExecutors(1)
handler.executorIdToContainer.keys.foreach { id => handler.killExecutor(id ) }
val statuses = Seq(container1, container2).map { c =>
ContainerStatus.newInstance(c.getId(), ContainerState.COMPLETE, "Finished", 0)
}
handler.updateResourceRequests()
handler.processCompletedContainers(statuses.toSeq)
handler.getNumExecutorsRunning should be (0)
handler.getNumPendingAllocate should be (1)
}
test("memory exceeded diagnostic regexes") {
val diagnostics =
"Container [pid=12465,containerID=container_1412887393566_0003_01_000002] is running " +
......
0% Loading or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment