[
https://issues.apache.org/jira/browse/SPARK-20107?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15981552#comment-15981552
]
Steve Loughran commented on SPARK-20107:
----------------------------------------
This does not solve the problem you think it does, not with S3.
Both commit algorithms assumes the filesystem is consistent, that a listFiles
will return the list of files it needs to rename. S3 isn't a consistent FS, so
newly created files aren't going to get picked up. It may appear to work well
in development, but in production there is a non-zero chance data will get
missed.
You need the HADOOP-13786 committer underneath spark for high performance,
fault tolerant renames.
> Add spark.hadoop.mapreduce.fileoutputcommitter.algorithm.version option to
> configuration.md
> -------------------------------------------------------------------------------------------
>
> Key: SPARK-20107
> URL: https://issues.apache.org/jira/browse/SPARK-20107
> Project: Spark
> Issue Type: Improvement
> Components: Documentation
> Affects Versions: 2.1.0
> Reporter: Yuming Wang
> Assignee: Yuming Wang
> Priority: Trivial
> Fix For: 2.2.0
>
>
> Set {{spark.hadoop.mapreduce.fileoutputcommitter.algorithm.version=2}} can
> speed up
> [HadoopMapReduceCommitProtocol#commitJob|https://github.com/apache/spark/blob/v2.1.0/core/src/main/scala/org/apache/spark/internal/io/HadoopMapReduceCommitProtocol.scala#L121]
> for many output files.
> It can speed up {{11 minutes}} for 216869 output files:
> {code:sql}
> CREATE TABLE tmp.spark_20107 AS SELECT
> category_id,
> product_id,
> track_id,
> concat(
> substr(ds, 3, 2),
> substr(ds, 6, 2),
> substr(ds, 9, 2)
> ) shortDate,
> CASE WHEN actiontype = '0' THEN 'browse' WHEN actiontype = '1' THEN 'fav'
> WHEN actiontype = '2' THEN 'cart' WHEN actiontype = '3' THEN 'order' ELSE
> 'invalid actio' END AS type
> FROM
> tmp.user_action
> WHERE
> ds > date_sub('2017-01-23', 730)
> AND actiontype IN ('0','1','2','3');
> {code}
> {code}
> $ hadoop fs -ls /user/hive/warehouse/tmp.db/spark_20107 | wc -l
> 216870
> {code}
> We should add this option to
> [configuration.md|http://spark.apache.org/docs/latest/configuration.html].
> All cloudera's hadoop 2.6.0-cdh5.4.0 or higher versions(see:
> [cloudera/hadoop-common@1c12361|https://github.com/cloudera/hadoop-common/commit/1c1236182304d4075276c00c4592358f428bc433]
> and
> [cloudera/hadoop-common@16b2de2|https://github.com/cloudera/hadoop-common/commit/16b2de27321db7ce2395c08baccfdec5562017f0])
> and apache's hadoop 2.7.0 or higher versions support this improvement.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
