[
https://issues.apache.org/jira/browse/HIVE-16257?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15961411#comment-15961411
]
Naveen Gangam commented on HIVE-16257:
--------------------------------------
[~xuefuz] In my case, it is always coming out to 4 tasks (in the first stage)
for some reason. There is a single input file on hdfs. However, the default
{{spark.executor.cores}} for me is 4. I dont know if they are related. After
adding some instrumentation, the data being passed to reducers was found to be
correct indicating the problem was prior to this phase. Seems like we use serde
to parse decimal types which pointed to HIVE-12768. It does not occur with HoMR
because each mapper is run in a seperate process. It also does not occur when
using a single executor.cores because it uses single thread to process it.
With the fix for HIVE-12768, we are currently unable to reproduce the issue
after several dozens of retries. As far as it being an intermittent issue, I
think the size of the data matters. The larger the dataset, the higher the
chance of reproducing it. In this case, the data was very small and hence was
hard to reproduce. Closing this jira. Thanks for all your help [~xuefuz]
> Intermittent issue with incorrect resultset with Spark
> ------------------------------------------------------
>
> Key: HIVE-16257
> URL: https://issues.apache.org/jira/browse/HIVE-16257
> Project: Hive
> Issue Type: Bug
> Components: Hive
> Affects Versions: 1.1.0
> Reporter: Naveen Gangam
> Attachments: failed_yarnlogs.log, success_yarnlogs.log
>
>
> This issue is highly intermittent that only seems to occurs with spark engine
> when the query has a GROUPBY clause. The following is the testcase.
> {code}
> drop table if exists test_hos_sample;
> create table test_hos_sample (name string, val1 decimal(18,2), val2
> decimal(20,3));
> insert into test_hos_sample values
> ('test1',101.12,102.123),('test1',101.12,102.123),('test2',102.12,103.234),('test1',101.12,102.123),('test3',103.52,102.345),('test3',103.52,102.345),('test3',103.52,102.345),('test3',103.52,102.345),('test3',103.52,102.345),('test4',104.52,104.456),('test4',104.52,104.456),('test5',105.52,105.567),('test3',103.52,102.345),('test5',105.52,105.567);
> set hive.execution.engine=spark;
> select name, val1,val2 from test_hos_sample group by name, val1, val2;
> {code}
> Expected Results:
> {code}
> name val1 val2
> test5 105.52 105.567
> test3 103.52 102.345
> test1 101.12 102.123
> test4 104.52 104.456
> test2 102.12 103.234
> {code}
> Incorrect results once in a while:
> {code}
> name val1 val2
> test5 105.52 105.567
> test3 103.52 102.345
> test1 104.52 102.123
> test4 104.52 104.456
> test2 102.12 103.234
> {code}
> 1) Not reproducible with HoMR.
> 2) Not an issue when running from spark-shell.
> 3) Not reproducible when the column data type is String or double. Only
> reproducible with decimal data types. Also works fine for decimal datatype if
> you cast decimal as string on read and cast it back to decimal on select.
> 4) Occurs with parquet and text file format as well. (havent tried with other
> formats).
> 5) Occurs in both scenarios when table data is within encryption zone and
> outside.
> 6) Even in clusters where this is reproducible, this occurs once in like 20
> times or more.
> 7) Occurs with both Beeline and Hive CLI.
> 8) Reproducible only when there is a a groupby clause.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
