holdenk commented on a change in pull request #29211:
URL: https://github.com/apache/spark/pull/29211#discussion_r466648846
##########
File path:
core/src/main/scala/org/apache/spark/executor/CoarseGrainedExecutorBackend.scala
##########
@@ -277,12 +282,59 @@ private[spark] class CoarseGrainedExecutorBackend(
if (executor != null) {
executor.decommission()
}
- logInfo("Done decommissioning self.")
+ // Shutdown the executor once all tasks are gone & any configured
migrations completed.
+ // Detecting migrations completion doesn't need to be perfect and we
want to minimize the
+ // overhead for executors that are not in decommissioning state as
overall that will be
+ // more of the executors. For example, this will not catch a block which
is already in
+ // the process of being put from a remote executor before migration
starts. This trade-off
+ // is viewed as acceptable to minimize introduction of any new locking
structures in critical
+ // code paths.
+
+ val shutdownThread = new Thread("wait-for-blocks-to-migrate") {
+ override def run(): Unit = {
+ var lastTaskRunningTime = System.nanoTime()
+ val sleep_time = 1000 // 1s
+
+ while (true) {
+ logInfo("Checking to see if we can shutdown.")
+ Thread.sleep(sleep_time)
+ if (executor == null || executor.numRunningTasks == 0) {
+ if (env.conf.get(STORAGE_DECOMMISSION_ENABLED)) {
+ logInfo("No running tasks, checking migrations")
+ val (migrationTime, allBlocksMigrated) =
env.blockManager.lastMigrationInfo()
+ // We can only trust allBlocksMigrated boolean value if there
were no tasks running
+ // since the start of computing it.
+ if (allBlocksMigrated && (migrationTime >
lastTaskRunningTime)) {
+ logInfo("No running tasks, all blocks migrated, stopping.")
+ exitExecutor(0, "Finished decommissioning", notifyDriver =
true)
+ } else {
+ logInfo("All blocks not yet migrated.")
+ }
+ } else {
+ logInfo("No running tasks, no block migration configured,
stopping.")
+ exitExecutor(0, "Finished decommissioning", notifyDriver =
true)
Review comment:
Unless it's an action. If someone just wants things to exit immediately
they should not enable decommissioning.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
