[
https://issues.apache.org/jira/browse/SPARK-5947?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14336556#comment-14336556
]
Philippe Girolami commented on SPARK-5947:
------------------------------------------
For some workloads, it can make more sense to use SKEWED ON rather than
PARTITION in order to prevent creating thousands of tiny partitions just to
handle a few large partitions.
As far as I can tell, these two cases can't be inferred from a directory layout
so maybe it would make sense to make PARTITION & SKEW part of Spark too, and
rely on meta-data defined by the application rather than directory discovery ?
> First class partitioning support in data sources API
> ----------------------------------------------------
>
> Key: SPARK-5947
> URL: https://issues.apache.org/jira/browse/SPARK-5947
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Reporter: Cheng Lian
>
> For file system based data sources, implementing Hive style partitioning
> support can be complex and error prone. To be specific, partitioning support
> include:
> # Partition discovery: Given a directory organized similar to Hive
> partitions, discover the directory structure and partitioning information
> automatically, including partition column names, data types, and values.
> # Reading from partitioned tables
> # Writing to partitioned tables
> It would be good to have first class partitioning support in the data sources
> API. For example, add a {{FileBasedScan}} trait with callbacks and default
> implementations for these features.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
