[jira] [Comment Edited] (HIVE-22098) Data loss occurs when multiple tables are join with different bucket_version

["JithendhiraKumar (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-22098?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17040749#comment-17040749 ] 

JithendhiraKumar edited comment on HIVE-22098 at 2/22/20 6:43 PM:
------------------------------------------------------------------

Hi, we are stuck with similar issue(scenario 2), after upgrading from 2.3 to 3.1.2. The tables which were created in Hive 2.3 has no bucketing version in its table properties, but when new tables are created in Hive 3.1.2 they are created with bucketing_version 2. When we do a left join between these old and new tables, in a few cases the results are wrong.

[~djaiswal] [~jdere]  Can you guys please review this patch?


was (Author: jithendhir92):
Hi, we are stuck with similar issue, after upgrading from 2.3 to 3.1.2. The tables which were created in Hive 2.3 has no bucketing version in its table properties, but when new tables are created in Hive 3.1.2 they are created with bucketing_version 2. When we do a left join between these old and new tables, in a few cases the results are wrong.

[~djaiswal] [~jdere]  Can you guys please review this patch?

> Data loss occurs when multiple tables are join with different bucket_version
> ----------------------------------------------------------------------------
>
>                 Key: HIVE-22098
>                 URL: https://issues.apache.org/jira/browse/HIVE-22098
>             Project: Hive
>          Issue Type: Bug
>          Components: Operators
>    Affects Versions: 3.1.0, 3.1.2
>            Reporter: LuGuangMing
>            Assignee: LuGuangMing
>            Priority: Blocker
>              Labels: data-loss, wrongresults
>         Attachments: HIVE-22098.1.patch, image-2019-08-12-18-45-15-771.png, join_test.sql, table_a_data.orc, table_b_data.orc, table_c_data.orc
>
>
> When different bucketVersion of tables do join and no of reducers number greater than 2, the result is incorrect (*data loss*).
>  *Scenario 1*: Three tables join. The temporary result data of table_a in the first table and table_b in the second table joins result is recorded as tmp_a_b, When it joins with the third table, the bucket_version=2 of the table created by default after hive-3.0.0, temporary data tmp_a_b initialized the bucketVerison=-1, and then ReduceSinkOperator Verketison=-1 is joined. In the init method, the hash algorithm of selecting join column is selected according to bucketVersion. If bucketVersion = 2 and is not an acid operation, it will acquired the new algorithm of hash. Otherwise, the old algorithm of hash is acquired. Because of the inconsistency of the algorithm of hash, the partition of data allocation caused are different. At stage of Reducer, Data with the same key can not be paired resulting in data loss.
> *Scenario 2*: create two test tables, create table table_bucketversion_1(col_1 string, col_2 string) TBLPROPERTIES ('bucketing_version'='1'); table_bucketversion_2(col_1 string, col_2 string) TBLPROPERTIES ('bucketing_version'='2');
>  when use table_bucketversion_1 to join table_bucketversion_2, partial result data will be loss due to bucketVerison is different.
>  



--
This message was sent by Atlassian Jira
(v8.3.4#803005)