> On Dec. 20, 2014, 1:36 a.m., Xuefu Zhang wrote:
> > itests/src/test/resources/testconfiguration.properties, line 545
> > <https://reviews.apache.org/r/29281/diff/1/?file=797796#file797796line545>
> >
> >     I think some of the "wrong" ordering is caused by different 
> > understanding of ordering. For instance, mapjoin1.q -> mapjoin10.q -> 
> > mapjoin11.q -> mapjoin12.q -> mapjoin2.q, or  mapjoin1.q -> mapjoin2.q -> 
> > mapjoin10.q -> mapjoin11.q -> mapjoin12.q. Since last reordering, we only 
> > added a few tests manually, which wouldn't have caused so much disrruption.
> 
> Szehon Ho wrote:
>     I ran Brock's script provided in the comments of 
> testconfiguration.properties that ensures the order.  It is based on unix 
> sort.  I thought it is the one we agreed to use?
> 
> Xuefu Zhang wrote:
>     Well, I think Jimmy used a slightly modified version of script to order 
> the list. He found Brock's script had certain issues. Did you see anything 
> obviouly wrong besides mapjoin1 -> mapjoin2 or mapjoin1 -> majoin10? It's up 
> to you either fix individual lines or reorder globally.

OK, I will keep this as its based on the current comment in 
testconfiguration.properties.  If Jimmy have a different scripts, let's put 
that script in the comments and regenerate the order in a separate JIRA.  Hope 
that is ok.


> On Dec. 20, 2014, 1:36 a.m., Xuefu Zhang wrote:
> > ql/src/java/org/apache/hadoop/hive/ql/optimizer/spark/SparkSMBJoinHintOptimizer.java,
> >  line 39
> > <https://reviews.apache.org/r/29281/diff/1/?file=797799#file797799line39>
> >
> >     The rule matches only MapJoinOperator, is it possible for nd a 
> > SMBMapJoinOperator?
> 
> Szehon Ho wrote:
>     It's a weird check that I copied from the mapreduce side.  It should not 
> happen in our use-case (or even their's).  I dont think it hurts though.  
> Should I get rid of it?
> 
> Xuefu Zhang wrote:
>     The check itself doesn't matter, but it causes confusion in 
> understanding. If you need to update the patch, let's remove it.

Done.


- Szehon


-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviews.apache.org/r/29281/#review65714
-----------------------------------------------------------


On Dec. 20, 2014, 3:32 a.m., Szehon Ho wrote:
> 
> -----------------------------------------------------------
> This is an automatically generated e-mail. To reply, visit:
> https://reviews.apache.org/r/29281/
> -----------------------------------------------------------
> 
> (Updated Dec. 20, 2014, 3:32 a.m.)
> 
> 
> Review request for hive.
> 
> 
> Bugs: HIVE-8640
>     https://issues.apache.org/jira/browse/HIVE-8640
> 
> 
> Repository: hive-git
> 
> 
> Description
> -------
> 
> This change is on the same principle as the refactoring of HIVE-8639.  The 
> goal is to move as much of the join optimization as possible to the same 
> traversal, and in fact the same process(joinOp) method, to simplify the logic 
> and also for compiler performance.
> 
> Whereas it is too hard to bring SparkMapJoinProcessor (for mapjoin hints) 
> into the same level due to the way it was written (see HIVE-8911), it is 
> possible to bring Bucket join and SMB join hints to the same level.  This 
> change introduces a parallel processor called 'SparkJoinHintOptimizer', which 
> takes a mapjoin already converted by SparkMapJoinProcessor as input and 
> converts it to Bucket/SMB join accordingly.  It runs alongside 
> 'SparkJoinOptimizer' which takes a common join operator and handles the 
> auto-conversion to mapjoin/bucketJoin/SMBJoin.
> 
> The one difference between mapjoin/bucketJoin vs SMB as Chao found was that 
> while Spark mapjoins expect RS for small-table branches in 
> mapjoin/bucketJoin, this is not expected for SMB join.  So I added a class 
> SparkSMBHintJoinOptimizer that first removes this before re-using the rest of 
> the existing code.
> 
> Another issue was found in NonBlockingOpDeDupProc that corrupts 
> 'mapJoinContext' data structure in the parse context.  A fix is offered in 
> HIVE-9117 and that should be committed to trunk and merged first, but it is 
> included here for reference.
> 
> 
> Diffs
> -----
> 
>   itests/src/test/resources/testconfiguration.properties fd732c1 
>   ql/src/java/org/apache/hadoop/hive/ql/optimizer/NonBlockingOpDeDupProc.java 
> 5e0959a 
>   
> ql/src/java/org/apache/hadoop/hive/ql/optimizer/spark/SparkJoinHintOptimizer.java
>  PRE-CREATION 
>   
> ql/src/java/org/apache/hadoop/hive/ql/optimizer/spark/SparkSMBJoinHintOptimizer.java
>  PRE-CREATION 
>   
> ql/src/java/org/apache/hadoop/hive/ql/optimizer/spark/SparkSortMergeJoinOptimizer.java
>  6a47513 
>   ql/src/java/org/apache/hadoop/hive/ql/parse/spark/SparkCompiler.java 
> 5227d92 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin1.q.out b18e02f 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin11.q.out bb7214c 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin12.q.out c0adef4 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin13.q.out 98d0706 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin2.q.out ea763c7 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin3.q.out 1b31561 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin4.q.out 97d2d74 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin5.q.out 94952a1 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin7.q.out ca59d02 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin8.q.out f419eaf 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin_negative2.q.out 
> b954feb 
>   ql/src/test/results/clientpositive/spark/bucketmapjoin_negative3.q.out 
> bfe5438 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin9.q.out d769ebe 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_1.q.out 8d0527e 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_10.q.out 2df87cf 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_11.q.out PRE-CREATION 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_12.q.out PRE-CREATION 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_13.q.out 5637206 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_14.q.out 3aed084 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_15.q.out 6ed680d 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_16.q.out a4fd7c3 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_17.q.out 6293450 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_2.q.out 1cf144b 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_3.q.out 6b44d2c 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_4.q.out d07d65a 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_5.q.out 607b1f0 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_6.q.out 30746ff 
>   ql/src/test/results/clientpositive/spark/smb_mapjoin_7.q.out c48ed6d 
> 
> Diff: https://reviews.apache.org/r/29281/diff/
> 
> 
> Testing
> -------
> 
> Re-enabled all the smb_mapjoin.* tests.
> 
> I saw that a lot of the tests are again not alphabetized, so re-ran the 
> script to alphabeticize them.  As part of that, realized that some tests like 
> 'bucket_map_join_spark.*' and 'join_empty' were missing proper comma 
> deliminters from the next test and probably not ran.  Also fixed the 
> windowing.q which is the last test.  This is all unrelated, but I am not sure 
> if they will trigger additional test failures if these were unintentionally 
> disabled.
> 
> 
> Thanks,
> 
> Szehon Ho
> 
>

Reply via email to