hive-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chao Sun (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HIVE-17018) Small table is converted to map join even the total size of small tables exceeds the threshold(hive.auto.convert.join.noconditionaltask.size)
Date Tue, 11 Jul 2017 06:20:00 GMT

    [ https://issues.apache.org/jira/browse/HIVE-17018?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16081743#comment-16081743
] 

Chao Sun commented on HIVE-17018:
---------------------------------

[~kellyzly]: yes, this is to prevent a Spark task from using unbounded memory for small tables.
I agree that the config {{hive.auto.convert.join.noconditionaltask.size}} is a little confusing.
It has already diverged from its original purpose for MR in several ways (e.g., Spark uses
in-memory data size while MR uses file size). A better way might be to have a separate config
just for HoS, and maybe a limit on small table memory per executor.

> Small table is converted to map join even the total size of small tables exceeds the
threshold(hive.auto.convert.join.noconditionaltask.size)
> ---------------------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: HIVE-17018
>                 URL: https://issues.apache.org/jira/browse/HIVE-17018
>             Project: Hive
>          Issue Type: Bug
>            Reporter: liyunzhang_intel
>            Assignee: liyunzhang_intel
>         Attachments: HIVE-17018_data_init.q, HIVE-17018.q, t3.txt
>
>
>  we use "hive.auto.convert.join.noconditionaltask.size" as the threshold. it means  the
sum of size for n-1 of the tables/partitions for a n-way join is smaller than it, it will
be converted to a map join. for example, A join B join C join D join E. Big table is A(100M),
small tables are B(10M),C(10M),D(10M),E(10M).  If we set hive.auto.convert.join.noconditionaltask.size=20M.
In current code, E,D,B will be converted to map join but C will not be converted to map join.
In my understanding, because hive.auto.convert.join.noconditionaltask.size can only contain
E and D, so C and B should not be converted to map join.  
> Let's explain more why E can be converted to map join.
> in current code, [SparkMapJoinOptimizer#getConnectedMapJoinSize|https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/optimizer/spark/SparkMapJoinOptimizer.java#L364]
calculates all the mapjoins  in the parent path and child path. The search stops when encountering
[UnionOperator or ReduceOperator|https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/optimizer/spark/SparkMapJoinOptimizer.java#L381].
Because C is not converted to map join because {{connectedMapJoinSize + totalSize) > maxSize}}
[see code|https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/optimizer/spark/SparkMapJoinOptimizer.java#L330].The
RS before the join of C remains. When calculating whether B will be converted to map join,
{{getConnectedMapJoinSize}} returns 0 as encountering [RS |https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/optimizer/spark/SparkMapJoinOptimizer.java#409]
and causes  {{connectedMapJoinSize + totalSize) < maxSize}} matches.
> [~xuefuz] or [~jxiang]: can you help see whether this is a bug or not  as you are more
familiar with SparkJoinOptimizer.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message