I was testing with spark yarn cluster mode.
The spark job runs in lower priority queue.
And its containers are preempted when a higher priority job comes.
However it relaunches the containers right after being killed.
And higher priority app kills them again.
So apps are stuck in this deadlock.
Infinite retry of executors is discussed here.
Found below trace in logs.
2019-05-20 03:40:07 [dispatcher-event-loop-0] INFO TaskSetManager :54 Task 95 failed because while it was being computed, its executor exited for a reason unrelated to the task. Not counting this failure towards the maximum number of failures for the task.
So it seems any retry count I set is not even considered.
Is there a flag to indicate that all failures in executor should be counted, and job should fail when maxFailures happen ?
spark version 2.11
Spark distinguishes between code throwing some exception and external issues, ie code failures and container failures.
But spark does not consider preemption as container failure.
See ApplicationMaster.scala, here spark decides to quit if container failure limit is hit.
It gets number of failed executors from YarnAllocator.
YarnAllocator updates its failed containers in some cases. But not for preemptions, see case ContainerExitStatus.PREEMPTED in same function.
We use spark 2.0.2, where code is slightly different but logic is same.
Fix seems to update failed containers collection for preemptions too.
Related
How can I increase failure tolerance on yarn? In a busy cluster my job fails due to too many failures. Most of the failures were due to Executor lost base by preemption.
If you have preemption enabled you really should be using the external shuffle service to avoid these issues. There's really not much that can be done otherwise.
https://issues.apache.org/jira/browse/SPARK-14209 - JIRA talks about.
Close yarn preemption?Or run smaller jobs to avoid complete recomputation?
i have run the job using spark-submit while that time we lost executor and the certain point we can recover or not if recover how we will recover and while how we have to get back that executor
You cannot handle executor failures programmatically in your application, if thats what you are asking.
You can configure spark configuration properties which guides the actual job execution including how YARN would schedule jobs and handle task and executor failures.
https://spark.apache.org/docs/latest/configuration.html#scheduling
Some important properties you may want to check out:
spark.task.maxFailures(default=4): Number of failures of any particular task
before giving up on the job. The total number of failures spread
across different tasks will not cause the job to fail; a particular
task has to fail this number of attempts. Should be greater than or
equal to 1. Number of allowed retries = this value - 1.
spark.blacklist.application.maxFailedExecutorsPerNode(default=2): (Experimental)
How many different executors must be blacklisted for the entire
application, before the node is blacklisted for the entire
application. Blacklisted nodes will be automatically added back to the
pool of available resources after the timeout specified by
spark.blacklist.timeout. Note that with dynamic allocation, though,
the executors on the node may get marked as idle and be reclaimed by
the cluster manager.
spark.blacklist.task.maxTaskAttemptsPerExecutor(default=1): (Experimental)
For a given task, how many times it can be retried on one executor
before the executor is blacklisted for that task.
I am seeing about 3018 tasks failed for the job as about 4 executors died.
The Executors summary (as below in Spark UI) have a completely different statistics. Out of 3018, about 2994 properly completed. My question is,
Will they be re-tried again?
Is there a config to override/limit this?
After monitoring the job and manually validation the attempt counts event for successful tasks, realised
Will they be re-tried again?
- Yes, even the successful tasks are retried.
Is there a config to override/limit this?
- Did not find any config to override this behaviour.
If an executer (kubernetes pod) dies (like with an OOM or timeout), all the tasks, even if successfully completed are re-executed. One of the main reason is, the shuffle writes from the executers are lost with the executor itself!!!
What's happening right now is YARN simply gets a number of executor from one spark job and give it to another spark job. As a result, this spark job encounters error and die.
Is there a way or an existing configuration where a certain spark job running on YARN have a fix resource allocation?
Fix resource allocation is an old concept and doesn't give benefit of proper resource utilization. Dynamic resource allocation is an advanced/expected feature of YARN. So, I recommend that you see what is happening actually. If a job is already running on then YARN doesn't take the resources and gives it to others. If resources are not available then the 2nd job will get queued and resources will not be pulled up abruptly from the 1st job. The reason is containers have a combination of memory and CPU. If memory is allocated to other job then basically it means that the JVM of the 1st job is lost for ever. YARN doesn't do what have mentioned.
We encounter a problem on a Spark job 1.6(on yarn) that never ends, whene several jobs launched simultaneously.
We found that by launching the job spark in yarn-client mode we do not have this problem, unlike launching it in yarn-cluster mode.
it could be a trail to find the cause.
we changed the code to add a sparkContext.stop ()
Indeed, the SparkContext was created (val sparkContext = createSparkContext) but not stopped. this solution has allowed us to decrease the number of jobs that remains blocked but nevertheless we still have some jobs blocked.
by analyzing the logs we have found this log that repeats without stopping:
17/09/29 11:04:37 DEBUG SparkEventPublisher: Enqueue SparkListenerExecutorMetricsUpdate(1,WrappedArray())
17/09/29 11:04:41 DEBUG ApplicationMaster: Sending progress
17/09/29 11:04:41 DEBUG ApplicationMaster: Number of pending allocations is 0. Sleeping for 5000.
it seems that the job block whene we call newAPIHadoopRDD to get data from Hbase. it may be the issue !!
Does someone have any idea about this issue ?
Thank you in advance