Hannu Kröger created SPARK-25597:
------------------------------------
Summary: SQL query with limit iterates the whole iterator when
WholeStage code generation is enabled
Key: SPARK-25597
URL: https://issues.apache.org/jira/browse/SPARK-25597
Project: Spark
Issue Type: Bug
Components: SQL
Affects Versions: 2.3.0
Reporter: Hannu Kröger
When _SELECT * FROM table LIMIT 1_ is executed, the WholeStageCodegenExec
generates code that iterates the source iterator completely instead of just
first element (per partition). This is a major performance penalty over the
expected functionality. This does not happen if _spark.sql.codegen.wholeStage_
is set to false in spark context conf.
In codegen mode {{LocalLimitExec}} overrides {{stopEarly}} method of
{{BufferedRowIterator}} to stop consuming rows when enough rows are fetched.
However, {{stopEarly}} is only called by code generated by {{InputAdapter}} and
{{InputAdapter}} is not added to the plan for some reason.
One potential way to fix this would be to check {{stopEarly}} in {{hasNext}}
method of {{BufferedRowIterator}}. Interesting fact is that the {{next}} does
not call {{hasNext}}. Therefore {{next}} might return empty result even if
there is data to be fetched by the {{processNext}} method. Or add a check to
stopEarly in shouldStop method.
This commit might have broken the original functionality:
[https://github.com/apache/spark/commit/b41ec997786e2be42a8a2a182212a610d08b221b]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
