[
https://issues.apache.org/jira/browse/SPARK-30511?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Zebing Lin updated SPARK-30511:
-------------------------------
Description:
*TL;DR*
When speculative tasks finished/failed/got killed, they are still considered
as pending and count towards the calculation of number of needed executors.
h3. Symptom
In one of our production job (where it's running 4 tasks per executor), we
found that it was holding 6 executors at the end with only 2 tasks running (1
speculative). With more logging enabled, we found the job printed:
{code:java}
pendingTasks is 0 pendingSpeculativeTasks is 17 totalRunningTasks is 2
{code}
while the job only had 1 speculative task running and 16 speculative tasks
intentionally killed because of corresponding original tasks had finished.
An easy repro of the issue (`--conf spark.speculation=true --conf
spark.executor.cores=4 --conf spark.dynamicAllocation.maxExecutors=1000` in
cluster mode):
{code:java}
val n = 4000
val someRDD = sc.parallelize(1 to n, n)
someRDD.mapPartitionsWithIndex( (index: Int, it: Iterator[Int]) => {
if (index < 300 && index >= 150) {
Thread.sleep(index * 1000) // Fake running tasks
} else if (index == 300) {
Thread.sleep(1000 * 1000) // Fake long running tasks
}
it.toList.map(x => index + ", " + x).iterator
}).collect
{code}
You will see when running the last task, we would be hold 39 executors:
!image-2020-01-15-11-09-29-215.png!
h3. The Bug
Upon examining the code of _pendingSpeculativeTasks_:
{code:java}
stageAttemptToNumSpeculativeTasks.map { case (stageAttempt, numTasks) =>
numTasks -
stageAttemptToSpeculativeTaskIndices.get(stageAttempt).map(_.size).getOrElse(0)
}.sum
{code}
where _stageAttemptToNumSpeculativeTasks(stageAttempt)_ is incremented on
_onSpeculativeTaskSubmitted_, but never decremented.
_stageAttemptToNumSpeculativeTasks -= stageAttempt_ is performed on stage
completion. *This means Spark is marking ended speculative tasks as pending,
which leads to Spark to hold more executors that it actually needs!*
I will have a PR ready to fix this issue, along with SPARK-2840 too
was:
*TL;DR*
When speculative tasks finished/failed/got killed, they are still considered
as pending and count towards the calculation of number of needed executors.
h3. Symptom
In one of our production job (where it's running 4 tasks per executor), we
found that it was holding 6 executors at the end with only 2 tasks running (1
speculative). With more logging enabled, we found the job printed:
{code:java}
pendingTasks is 0 pendingSpeculativeTasks is 17 totalRunningTasks is 2
{code}
while the job only had 1 speculative task running and 16 speculative tasks
intentionally killed because of corresponding original tasks had finished.
h3. The Bug
Upon examining the code of _pendingSpeculativeTasks_:
{code:java}
stageAttemptToNumSpeculativeTasks.map { case (stageAttempt, numTasks) =>
numTasks -
stageAttemptToSpeculativeTaskIndices.get(stageAttempt).map(_.size).getOrElse(0)
}.sum
{code}
where _stageAttemptToNumSpeculativeTasks(stageAttempt)_ is incremented on
_onSpeculativeTaskSubmitted_, but never decremented.
_stageAttemptToNumSpeculativeTasks -= stageAttempt_ is performed on stage
completion. *This means Spark is marking ended speculative tasks as pending,
which leads to Spark to hold more executors that it actually needs!*
I will have a PR ready to fix this issue, along with SPARK-2840 too
> Spark marks ended speculative tasks as pending leads to holding idle executors
> ------------------------------------------------------------------------------
>
> Key: SPARK-30511
> URL: https://issues.apache.org/jira/browse/SPARK-30511
> Project: Spark
> Issue Type: Bug
> Components: Scheduler
> Affects Versions: 2.3.0
> Reporter: Zebing Lin
> Priority: Major
>
> *TL;DR*
> When speculative tasks finished/failed/got killed, they are still considered
> as pending and count towards the calculation of number of needed executors.
> h3. Symptom
> In one of our production job (where it's running 4 tasks per executor), we
> found that it was holding 6 executors at the end with only 2 tasks running (1
> speculative). With more logging enabled, we found the job printed:
> {code:java}
> pendingTasks is 0 pendingSpeculativeTasks is 17 totalRunningTasks is 2
> {code}
> while the job only had 1 speculative task running and 16 speculative tasks
> intentionally killed because of corresponding original tasks had finished.
> An easy repro of the issue (`--conf spark.speculation=true --conf
> spark.executor.cores=4 --conf spark.dynamicAllocation.maxExecutors=1000` in
> cluster mode):
> {code:java}
> val n = 4000
> val someRDD = sc.parallelize(1 to n, n)
> someRDD.mapPartitionsWithIndex( (index: Int, it: Iterator[Int]) => {
> if (index < 300 && index >= 150) {
> Thread.sleep(index * 1000) // Fake running tasks
> } else if (index == 300) {
> Thread.sleep(1000 * 1000) // Fake long running tasks
> }
> it.toList.map(x => index + ", " + x).iterator
> }).collect
> {code}
> You will see when running the last task, we would be hold 39 executors:
> !image-2020-01-15-11-09-29-215.png!
> h3. The Bug
> Upon examining the code of _pendingSpeculativeTasks_:
> {code:java}
> stageAttemptToNumSpeculativeTasks.map { case (stageAttempt, numTasks) =>
> numTasks -
> stageAttemptToSpeculativeTaskIndices.get(stageAttempt).map(_.size).getOrElse(0)
> }.sum
> {code}
> where _stageAttemptToNumSpeculativeTasks(stageAttempt)_ is incremented on
> _onSpeculativeTaskSubmitted_, but never decremented.
> _stageAttemptToNumSpeculativeTasks -= stageAttempt_ is performed on stage
> completion. *This means Spark is marking ended speculative tasks as pending,
> which leads to Spark to hold more executors that it actually needs!*
> I will have a PR ready to fix this issue, along with SPARK-2840 too
>
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
