[
https://issues.apache.org/jira/browse/SPARK-21595?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16114255#comment-16114255
]
Stephan Reiling edited comment on SPARK-21595 at 8/4/17 11:19 AM:
------------------------------------------------------------------
I have tried out a couple of settings for
spark.sql.windowExec.buffer.spill.threshold and I have now settled on 4M as the
default for it in my workflows. This gives about the same behavior as spark
2.1. But this is dependent on the amount of spark memory and the size of the
rows in the dataframe.
I am not in favor of introducing another threshold for this. If the spilling is
delayed, but then happens with the low threshold of 4096 rows, in my case this
would still spill 110k files to disk and potentially cause a "too many open
files" exception (right ?).
Just looking at the spilling behavior, it would be better if the value would
not specify the number of rows, but the amount of memory. So instead of 4096
rows, it would specify 500MB of memory, and then spill chunks of 500MB to disk.
How many rows this is would change case by case.
was (Author: sreiling):
I have tried out a couple of settings for
spark.sql.windowExec.buffer.spill.threshold and I have now settled on 4M as the
default for it in my work flows. This gives about the same behavior as spark
2.1. But this is dependent on the amount of spark memory and the size of the
rows in the dataframe.
I am not in favor of introducing another threshold for this. If the spilling is
delayed, but then happens with the low threshold of 4096 rows, in my case this
would still spill 110k files to disk and potentially cause a "too many open
files" exception (right ?).
Just looking at the spilling behavior, it would be better if the value would
not specify the number of rows, but the amount of memory. So instead of 4096
rows, it would specify 500MB of memory, and then spill chunks of 500MB to disk.
How many rows this is would change case by case.
> introduction of spark.sql.windowExec.buffer.spill.threshold in spark 2.2
> breaks existing workflow
> -------------------------------------------------------------------------------------------------
>
> Key: SPARK-21595
> URL: https://issues.apache.org/jira/browse/SPARK-21595
> Project: Spark
> Issue Type: Bug
> Components: Documentation, PySpark
> Affects Versions: 2.2.0
> Environment: pyspark on linux
> Reporter: Stephan Reiling
> Priority: Minor
> Labels: documentation, regression
>
> My pyspark code has the following statement:
> {code:java}
> # assign row key for tracking
> df = df.withColumn(
> 'association_idx',
> sqlf.row_number().over(
> Window.orderBy('uid1', 'uid2')
> )
> )
> {code}
> where df is a long, skinny (450M rows, 10 columns) dataframe. So this creates
> one large window for the whole dataframe to sort over.
> In spark 2.1 this works without problem, in spark 2.2 this fails either with
> out of memory exception or too many open files exception, depending on memory
> settings (which is what I tried first to fix this).
> Monitoring the blockmgr, I see that spark 2.1 creates 152 files, spark 2.2
> creates >110,000 files.
> In the log I see the following messages (110,000 of these):
> {noformat}
> 17/08/01 08:55:37 INFO UnsafeExternalSorter: Spilling data because number of
> spilledRecords crossed the threshold 4096
> 17/08/01 08:55:37 INFO UnsafeExternalSorter: Thread 156 spilling sort data of
> 64.1 MB to disk (0 time so far)
> 17/08/01 08:55:37 INFO UnsafeExternalSorter: Spilling data because number of
> spilledRecords crossed the threshold 4096
> 17/08/01 08:55:37 INFO UnsafeExternalSorter: Thread 156 spilling sort data of
> 64.1 MB to disk (1 time so far)
> {noformat}
> So I started hunting for clues in UnsafeExternalSorter, without luck. What I
> had missed was this one message:
> {noformat}
> 17/08/01 08:55:37 INFO ExternalAppendOnlyUnsafeRowArray: Reached spill
> threshold of 4096 rows, switching to
> org.apache.spark.util.collection.unsafe.sort.UnsafeExternalSorter
> {noformat}
> Which allowed me to track down the issue.
> By changing the configuration to include:
> {code:java}
> spark.sql.windowExec.buffer.spill.threshold 2097152
> {code}
> I got it to work again and with the same performance as spark 2.1.
> I have workflows where I use windowing functions that do not fail, but took a
> performance hit due to the excessive spilling when using the default of 4096.
> I think to make it easier to track down these issues this config variable
> should be included in the configuration documentation.
> Maybe 4096 is too small of a default value?
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
