[
https://issues.apache.org/jira/browse/SPARK-21056?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17950467#comment-17950467
]
Ohad Raviv commented on SPARK-21056:
------------------------------------
Hi! I know it's been a while :D, but we've recently discovered significant
resource waste related to this issue. The root cause appears to be the folder
structure in S3 (`{{{}/year/month/day/hour`{}}}), which led the driver to
perform repeated serial listings over an extended period. This resulted in
prolonged job durations and inefficient use of resources.
We found that adjusting the `{{{}threshold`{}}} parameter to offload some of
the listing work to executors does help. However, there’s still considerable
back-and-forth between the driver and executors, as the driver's serial loops
continue to generate a large number of jobs.
Any idea why this solution wasn't accepted?
> InMemoryFileIndex.listLeafFiles should create at most one spark job when
> listing files in parallel
> --------------------------------------------------------------------------------------------------
>
> Key: SPARK-21056
> URL: https://issues.apache.org/jira/browse/SPARK-21056
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 2.1.1, 2.2.0
> Reporter: Bertrand Bossy
> Priority: Major
> Labels: bulk-closed
>
> Given partitioned file relation (e.g. parquet):
> {code}
> root/a=../b=../c=..
> {code}
> InMemoryFileIndex.listLeafFiles runs numberOfPartitions(a) times
> numberOfPartitions(b) spark jobs sequentially to list leaf files, if both
> numberOfPartitions(a) and numberOfPartitions(b) are below
> {{spark.sql.sources.parallelPartitionDiscovery.threshold}} and
> numberOfPartitions(c) is above
> {{spark.sql.sources.parallelPartitionDiscovery.threshold}}
> Since the jobs are run sequentially, the overhead of the jobs dominates and
> the file listing operation can become significantly slower than listing the
> files from the driver.
> I propose that InMemoryFileIndex.listLeafFiles should launch at most one
> spark job for listing leaf files.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
