[
https://issues.apache.org/jira/browse/HIVE-2206?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13554712#comment-13554712
]
Ashutosh Chauhan commented on HIVE-2206:
----------------------------------------
I did some testing of this on our use-cases. Lets say you have two tables:
{code}
create table t1 (a int, b int);
create table t2 (c int, d int);
select a from ( select * from t1 join t2 on (t1.a = t2.c) )e group by a;
select a from ( select/*+ MAPJOIN(t2) */ * from t1 join t2 on (t1.a = t2.c) )e
group by a;
{code}
Now, ysmart is able to optimize first query fine, where it fuses 2MR jobs in 1
MR, since join and group-by has same key.
However this doesn't work with 2nd query which has mapjoin. This results in 2
MR job. This is especially important since in map-join case you don't need the
condition of join-key being same as groupby key, which is *very* important. In
our use-cases, we have observed its rarely the case that join and group-by is
on same key. But, in most cases we are able to utilize map-join, since data
which we are joining on is small enough. And than subsequent group-by which is
on different key can happen on the reduce side of this single MR job.
Any thoughts on how this could be achieved ?
Though, it looks like for this use-case we don't need ysmart since we don't
need to detect any correlation. We can walk on query plan and can always fuse
map-side join followed by group-by (i.e., no need to detect any correlation
among inputs or to detect whether join and group-by keys are same).
> add a new optimizer for query correlation discovery and optimization
> --------------------------------------------------------------------
>
> Key: HIVE-2206
> URL: https://issues.apache.org/jira/browse/HIVE-2206
> Project: Hive
> Issue Type: New Feature
> Components: Query Processor
> Affects Versions: 0.10.0
> Reporter: He Yongqiang
> Assignee: Yin Huai
> Attachments: HIVE-2206.10-r1384442.patch.txt,
> HIVE-2206.11-r1385084.patch.txt, HIVE-2206.12-r1386996.patch.txt,
> HIVE-2206.13-r1389072.patch.txt, HIVE-2206.14-r1389704.patch.txt,
> HIVE-2206.15-r1392491.patch.txt, HIVE-2206.16-r1399936.patch.txt,
> HIVE-2206.17-r1404933.patch.txt, HIVE-2206.18-r1407720.patch.txt,
> HIVE-2206.19-r1410581.patch.txt, HIVE-2206.1.patch.txt,
> HIVE-2206.2.patch.txt, HIVE-2206.3.patch.txt, HIVE-2206.4.patch.txt,
> HIVE-2206.5-1.patch.txt, HIVE-2206.5.patch.txt, HIVE-2206.6.patch.txt,
> HIVE-2206.7.patch.txt, HIVE-2206.8.r1224646.patch.txt,
> HIVE-2206.8-r1237253.patch.txt, testQueries.2.q, YSmartPatchForHive.patch
>
>
> This issue proposes a new logical optimizer called Correlation Optimizer,
> which is used to merge correlated MapReduce jobs (MR jobs) into a single MR
> job. The idea is based on YSmart (http://ysmart.cse.ohio-state.edu/).The
> paper and slides of YSmart are linked at the bottom.
> Since Hive translates queries in a sentence by sentence fashion, for every
> operation which may need to shuffle the data (e.g. join and aggregation
> operations), Hive will generate a MapReduce job for that operation. However,
> for those operations which may need to shuffle the data, they may involve
> correlations explained below and thus can be executed in a single MR job.
> # Input Correlation: Multiple MR jobs have input correlation (IC) if their
> input relation sets are not disjoint;
> # Transit Correlation: Multiple MR jobs have transit correlation (TC) if they
> have not only input correlation, but also the same partition key;
> # Job Flow Correlation: An MR has job flow correlation (JFC) with one of its
> child nodes if it has the same partition key as that child node.
> The current implementation of correlation optimizer only detect correlations
> among MR jobs for reduce-side join operators and reduce-side aggregation
> operators (not map only aggregation). A query will be optimized if it
> satisfies following conditions.
> # There exists a MR job for reduce-side join operator or reduce side
> aggregation operator which have JFC with all of its parents MR jobs (TCs will
> be also exploited if JFC exists);
> # All input tables of those correlated MR job are original input tables (not
> intermediate tables generated by sub-queries); and
> # No self join is involved in those correlated MR jobs.
> Correlation optimizer is implemented as a logical optimizer. The main reasons
> are that it only needs to manipulate the query plan tree and it can leverage
> the existing component on generating MR jobs.
> Current implementation can serve as a framework for correlation related
> optimizations. I think that it is better than adding individual optimizers.
> There are several work that can be done in future to improve this optimizer.
> Here are three examples.
> # Support queries only involve TC;
> # Support queries in which input tables of correlated MR jobs involves
> intermediate tables; and
> # Optimize queries involving self join.
> References:
> Paper and presentation of YSmart.
> Paper:
> http://www.cse.ohio-state.edu/hpcs/WWW/HTML/publications/papers/TR-11-7.pdf
> Slides: http://sdrv.ms/UpwJJc
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
