[
https://issues.apache.org/jira/browse/SPARK-23963?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16442783#comment-16442783
]
Ruslan Dautkhanov commented on SPARK-23963:
-------------------------------------------
I thought it's just a matter of a Spark committer to commit the same PR
[https://github.com/apache/spark/pull/21043] to a different branch? Spark2.2 in
this case.
This PR gives 24x improvement on 6000 columns as you discovered, so I think
this 1-line change should be admitted to Spark 2.2 fairly easily.
> Queries on text-based Hive tables grow disproportionately slower as the
> number of columns increase
> --------------------------------------------------------------------------------------------------
>
> Key: SPARK-23963
> URL: https://issues.apache.org/jira/browse/SPARK-23963
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 2.3.0
> Reporter: Bruce Robbins
> Assignee: Bruce Robbins
> Priority: Minor
> Fix For: 2.4.0
>
>
> TableReader gets disproportionately slower as the number of columns in the
> query increase.
> For example, reading a table with 6000 columns is 4 times more expensive per
> record than reading a table with 3000 columns, rather than twice as expensive.
> The increase in processing time is due to several Lists (fieldRefs,
> fieldOrdinals, and unwrappers), each of which the reader accesses by column
> number for each column in a record. Because each List has O\(n\) time for
> lookup by column number, these lookups grow increasingly expensive as the
> column count increases.
> When I patched the code to change those 3 Lists to Arrays, the query times
> became proportional.
>
>
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
