[
https://issues.apache.org/jira/browse/HIVE-21915?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16872385#comment-16872385
]
Wei Zhang commented on HIVE-21915:
----------------------------------
UPDATE:
We have to set hive.merge.tezfiles=true; to reproduce this issue, and updated
the test case to turn on file merge.
In our settings, hive.merge.tezfiles defaults to true. Ignored this factor
before.
> Hive with TEZ UNION ALL and UDTF results in data loss
> -----------------------------------------------------
>
> Key: HIVE-21915
> URL: https://issues.apache.org/jira/browse/HIVE-21915
> Project: Hive
> Issue Type: Bug
> Components: Query Processor
> Affects Versions: 1.2.1
> Reporter: Wei Zhang
> Assignee: Wei Zhang
> Priority: Major
> Attachments: HIVE-21915.01.patch, HIVE-21915.02.patch,
> HIVE-21915.03.patch
>
>
> The HQL syntax is like this:
> CREATE TEMPORARY TABLE tez_union_all_loss_data AS
> SELECT xxx, yyy, zzz,1 as tag
> FROM ods_1
> UNION ALL
> SELECT xxx, yyy, zzz, tag
> FROM
> (
> SELECT xxx
> ,get_json_object(get_json_object(tb,'$.a'),'$.b') AS yyy
> ,zzz
> ,2 as tag
> FROM ods_2
> LATERAL VIEW EXPLODE(some_udf(uuu)) team_number AS tb
> ) tbl
> ;
>
> With above HQL, we are expecting that rows with both tag = 2 and tag = 1
> appear. In our case however, all the rows with tag = 1 are lost.
> Dig deeper we can find that the generated two maps have identical task tmp
> paths. And that results from when UDTF is present, the FileSinkOperator will
> be processed twice generating the tmp path in
> GenTezUtils.removeUnionOperators();
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
