[
https://issues.apache.org/jira/browse/HIVE-11954?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14955449#comment-14955449
]
Laljo John Pullokkaran commented on HIVE-11954:
-----------------------------------------------
#1 If a relation has input size > maxSize (i.e it can not be kept in memory)
then we shouldn't use costlyops comparison. Instead that relation should be
chosen as streaming side.
The check in Line 581 could prevent an input that has lower costlyops but whose
size > maxis from becoming streaming side.
if (bigInputNumberCostlyOps == -1 || inputNumberCostlyOps >
bigInputNumberCostlyOps
|| (inputNumberCostlyOps == bigInputNumberCostlyOps && inputSize
> bigInputStat.getDataSize())) {
> Extend logic to choose side table in MapJoin Conversion algorithm
> -----------------------------------------------------------------
>
> Key: HIVE-11954
> URL: https://issues.apache.org/jira/browse/HIVE-11954
> Project: Hive
> Issue Type: Bug
> Components: Physical Optimizer
> Affects Versions: 2.0.0
> Reporter: Jesus Camacho Rodriguez
> Assignee: Jesus Camacho Rodriguez
> Attachments: HIVE-11954.01.patch, HIVE-11954.02.patch,
> HIVE-11954.03.patch, HIVE-11954.04.patch, HIVE-11954.05.patch,
> HIVE-11954.06.patch, HIVE-11954.patch, HIVE-11954.patch
>
>
> Selection of side table (in memory/hash table) in MapJoin Conversion
> algorithm needs to be more sophisticated.
> In an N way Map Join, Hive should pick an input stream as side table (in
> memory table) that has least cost in producing relation (like TS(FIL|Proj)*).
> Cost based choice needs extended cost model; without return path its going to
> be hard to do this.
> For the time being we could employ a modified cost based algorithm for side
> table selection.
> New algorithm is described below:
> 1. Identify the candidate set of inputs for side table (in memory/hash table)
> from the inputs (based on conditional task size)
> 2. For each of the input identify its cost, memory requirement. Cost is 1 for
> each heavy weight relation op (Join, GB, PTF/Windowing, TF, etc.). Cost for
> an input is the total no of heavy weight ops in its branch.
> 3. Order set from #1 on cost & memory req (ascending order)
> 4. Pick the first element from #3 as the side table.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
