[HIVE-23891] UNION ALL and multiple task attempts can cause file duplication

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.0.0-alpha-2
• Labels: pull-request-available
• Assignee: Zhihua Deng
• Reporter: George Pachitariu
• Created: 21/Jul/20 17:07
• Updated: 10/Jul/23 07:51
• Resolved: 24/Jan/23 07:20

Description

Hello, 

the specific scenario when this can happen:

• the execution engine is Tez;
• speculative execution is on;
• the query inserts into a table and the last step is a UNION sql clause;

The problem is that Tez creates an extra layer of subdirectories when there is a UNION. Later, when deduplicating, Hive doesn't take that into account and only deduplicates folders but not the files inside.

So for a query like this:

insert overwrite table union_all
    select * from union_first_part
union all
    select * from union_second_part;

The folder structure afterwards will be like this (a possible example):

.../union_all/HIVE_UNION_SUBDIR_1/000000_0
.../union_all/HIVE_UNION_SUBDIR_1/000000_1
.../union_all/HIVE_UNION_SUBDIR_2/000000_1

The attached patch increases the number of folder levels that Hive will check recursively for duplicates when we have a UNION in Tez.

Feel free to reach out if you have any questions .

Issue Links

• is related to: HIVE-27494 Deduplicate the task result that generated by more branches in union all
• links to: GitHub Pull Request #1294
• links to: GitHub Pull Request #3836