spark-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Kay Ousterhout (JIRA)" <>
Subject [jira] [Commented] (SPARK-19326) Speculated task attempts do not get launched in few scenarios
Date Fri, 03 Feb 2017 20:34:51 GMT


Kay Ousterhout commented on SPARK-19326:

I see that makes sense; thanks for the additional explanation.  [~andrewor14] did you think
about this issue when implementing dynamic allocation originally? I noticed there'a a [comment
saying that speculation is not considered for simplicity](,
but it does seem like this functionality can prevent speculation from occurring.

> Speculated task attempts do not get launched in few scenarios
> -------------------------------------------------------------
>                 Key: SPARK-19326
>                 URL:
>             Project: Spark
>          Issue Type: Bug
>          Components: Scheduler
>    Affects Versions: 2.0.2, 2.1.0
>            Reporter: Tejas Patil
> Speculated copies of tasks do not get launched in some cases.
> Examples:
> - All the running executors have no CPU slots left to accommodate a speculated copy of
the task(s). If the all running executors reside over a set of slow / bad hosts, they will
keep the job running for long time
> - `spark.task.cpus` > 1 and the running executor has not filled up all its CPU slots.
Since the [speculated copies of tasks should run on different host|]
and not the host where the first copy was launched.
> In both these cases, `ExecutorAllocationManager` does not know about pending speculation
task attempts and thinks that all the resource demands are well taken care of. ([relevant
> This adds variation in the job completion times and more importantly SLA misses :( In
prod, with a large number of jobs, I see this happening more often than one would think. Chasing
the bad hosts or reason for slowness doesn't scale.
> Here is a tiny repro. Note that you need to launch this with (Mesos or YARN or standalone
deploy mode) along with `--conf spark.speculation=true --conf spark.executor.cores=4 --conf
> {code}
> val n = 100
> val someRDD = sc.parallelize(1 to n, n)
> someRDD.mapPartitionsWithIndex( (index: Int, it: Iterator[Int]) => {
> if (index == 1) {
>   Thread.sleep(Long.MaxValue)  // fake long running task(s)
> }
> => index + ", " + x).iterator
> }).collect
> {code}

This message was sent by Atlassian JIRA

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message