[
https://issues.apache.org/jira/browse/SPARK-20227?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15960953#comment-15960953
]
Quentin Auge edited comment on SPARK-20227 at 4/7/17 3:26 PM:
--------------------------------------------------------------
Well, you were right to ask. After further investigation, it seems the job does
not hang forever, but takes much longer to finish when n is large enough.
I made a mistake when I mentioned the job finishes in a sensible amount of time
for n = 20. Really, it starts to hang badly from n = 12 (so 11 successive
aggregations).
Please see the following chart:
!https://image.ibb.co/fxRL85/figure_1.png!
n on x axis, time for the job to complete in seconds on y axis.
The job takes as much as 1 hour and 47 minutes to complete for n = 14.
For the most part, the amount of time the job takes to complete from n = 12 is
spent after this message:
{code}
17/04/05 14:39:28 INFO ExecutorAllocationManager: Removing executor 1 because
it has been idle for 60 seconds (new desired total will be 0)
17/04/05 14:39:29 INFO YarnSchedulerBackend$YarnDriverEndpoint: Disabling
executor 1.
17/04/05 14:39:29 INFO DAGScheduler: Executor lost: 1 (epoch 0)
17/04/05 14:39:29 INFO BlockManagerMasterEndpoint: Trying to remove executor 1
from BlockManagerMaster.
17/04/05 14:39:29 INFO BlockManagerMasterEndpoint: Removing block manager
BlockManagerId(1, ip-172-30-0-149.ec2.internal, 35666, None)
17/04/05 14:39:29 INFO BlockManagerMaster: Removed 1 successfully in
removeExecutor
17/04/05 14:39:29 INFO YarnScheduler: Executor 1 on
ip-172-30-0-149.ec2.internal killed by driver.
17/04/05 14:39:29 INFO ExecutorAllocationManager: Existing executor 1 has been
removed (new total is 0)
{code}
Except for the master, the cluster remains idle. Right after that, the job
starts again and completes.
was (Author: quentin):
Well, you were right to ask. After further investigation, it seems the job does
not hang forever, but takes much longer to finish when n is large enough.
I made a mistake when I mentioned the job finishes in a sensible amount of time
for n = 20. Really, it starts to hang badly from n = 12 (so 11 successive
aggregations).
See the following figure:
!https://image.ibb.co/fxRL85/figure_1.png!
n on x axis, time for the job to complete in seconds on y axis.
The job takes as much as 1 hour and 47 minutes to complete for n = 14.
For the most part, the amount of time the job takes to complete from n = 12 is
spent after this message:
{code}
17/04/05 14:39:28 INFO ExecutorAllocationManager: Removing executor 1 because
it has been idle for 60 seconds (new desired total will be 0)
17/04/05 14:39:29 INFO YarnSchedulerBackend$YarnDriverEndpoint: Disabling
executor 1.
17/04/05 14:39:29 INFO DAGScheduler: Executor lost: 1 (epoch 0)
17/04/05 14:39:29 INFO BlockManagerMasterEndpoint: Trying to remove executor 1
from BlockManagerMaster.
17/04/05 14:39:29 INFO BlockManagerMasterEndpoint: Removing block manager
BlockManagerId(1, ip-172-30-0-149.ec2.internal, 35666, None)
17/04/05 14:39:29 INFO BlockManagerMaster: Removed 1 successfully in
removeExecutor
17/04/05 14:39:29 INFO YarnScheduler: Executor 1 on
ip-172-30-0-149.ec2.internal killed by driver.
17/04/05 14:39:29 INFO ExecutorAllocationManager: Existing executor 1 has been
removed (new total is 0)
{code}
Except for the master, the cluster remains idle. Right after that, the job
starts again and completes.
> Job hangs when joining a lot of aggregated columns
> --------------------------------------------------
>
> Key: SPARK-20227
> URL: https://issues.apache.org/jira/browse/SPARK-20227
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 2.1.0
> Environment: AWS emr-5.4.0, master: m4.xlarge, core: 4 m4.xlarge
> Reporter: Quentin Auge
>
> I'm trying to replace a lot of different columns in a dataframe with
> aggregates of themselves, and then join the resulting dataframe.
> {code}
> # Create a dataframe with 1 row and 50 columns
> n = 50
> df = sc.parallelize([Row(*range(n))]).toDF()
> cols = df.columns
> # Replace each column values with aggregated values
> window = Window.partitionBy(cols[0])
> for col in cols[1:]:
> df = df.withColumn(col, sum(col).over(window))
> # Join
> other_df = sc.parallelize([Row(0)]).toDF()
> result = other_df.join(df, on = cols[0])
> result.show()
> {code}
> Spark hangs forever when executing the last line. The strange thing is, it
> depends on the number of columns. Spark does not hang for n = 5, 10, or 20
> columns. For n = 50 and beyond, it does.
> {code}
> 17/04/05 14:39:28 INFO ExecutorAllocationManager: Removing executor 1 because
> it has been idle for 60 seconds (new desired total will be 0)
> 17/04/05 14:39:29 INFO YarnSchedulerBackend$YarnDriverEndpoint: Disabling
> executor 1.
> 17/04/05 14:39:29 INFO DAGScheduler: Executor lost: 1 (epoch 0)
> 17/04/05 14:39:29 INFO BlockManagerMasterEndpoint: Trying to remove executor
> 1 from BlockManagerMaster.
> 17/04/05 14:39:29 INFO BlockManagerMasterEndpoint: Removing block manager
> BlockManagerId(1, ip-172-30-0-149.ec2.internal, 35666, None)
> 17/04/05 14:39:29 INFO BlockManagerMaster: Removed 1 successfully in
> removeExecutor
> 17/04/05 14:39:29 INFO YarnScheduler: Executor 1 on
> ip-172-30-0-149.ec2.internal killed by driver.
> 17/04/05 14:39:29 INFO ExecutorAllocationManager: Existing executor 1 has
> been removed (new total is 0)
> {code}
> All executors are inactive and thus killed after 60 seconds, the master
> spends some CPU on a process that hangs indefinitely, and the workers are
> idle.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
