[
https://issues.apache.org/jira/browse/ASTERIXDB-2784?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17203463#comment-17203463
]
Shiva Jahangiri commented on ASTERIXDB-2784:
--------------------------------------------
The fix for this issue is in my next patch for hybrid hash join where I define
NoGrow-NoSteal and Grow-Steal policies. Feel free to assign it to me.
> Join memory requirement for large objects
> -----------------------------------------
>
> Key: ASTERIXDB-2784
> URL: https://issues.apache.org/jira/browse/ASTERIXDB-2784
> Project: Apache AsterixDB
> Issue Type: Improvement
> Components: COMP - Compiler, RT - Runtime
> Reporter: Chen Luo
> Priority: Major
>
> Currently the compiler assumes the minimum number of join frames is 5 [1].
> However, this does not guarantee a join will always succeed in case of large
> objects. The actual join memory requirement is actually MAX(5, #partitions *
> #large object size). The reason is that in the spill policy [2], we only
> spill a partition if it hasn't been spilled before. As a result, when we are
> writing to an empty partition, it is possible that each of other partitions
> has one large object (which could be larger than the frame size) but no
> partition can be spilled. Thus, the join memory requirement becomes
> #partitions * #large object size in this case.
> [1]
> [https://github.com/apache/asterixdb/blob/master/hyracks-fullstack/algebricks/algebricks-core/src/main/java/org/apache/hyracks/algebricks/core/algebra/operators/physical/AbstractJoinPOperator.java#L29)|https://github.com/apache/asterixdb/blob/master/hyracks-fullstack/algebricks/algebricks-core/src/main/java/org/apache/hyracks/algebricks/core/algebra/operators/physical/AbstractJoinPOperator.java#L29).]
> [2]
> https://github.com/apache/asterixdb/blob/37dfed60fb47afcc86de6d17704a8f100217057d/hyracks-fullstack/hyracks/hyracks-dataflow-std/src/main/java/org/apache/hyracks/dataflow/std/buffermanager/PreferToSpillFullyOccupiedFramePolicy.java#L55
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
