[HIVE-26009] Determine number of buckets for implicitly bucketed ACIDv2 tables

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Labels: pull-request-available
• Assignee: Simhadri Govindappa
• Reporter: Simhadri Govindappa
• Created: 07/Mar/22 09:46
• Updated: 09/May/22 22:10

Description

Hive tries to set number of reducers equal to number of buckets here: 
https://github.com/apache/hive/blob/9857c4e584384f7b0a49c34bc2bdf876c2ea1503/ql/src/java/org/apache/hadoop/hive/ql/parse/SemanticAnalyzer.java#L6958 

 

The numberOfBuckets for implicitly bucketed tables is set to -1 by default. When this is the case, it is left to hive to estimate the number of reducers required the job, based on job input, and configuration parameters.

https://github.com/apache/hive/blob/9857c4e584384f7b0a49c34bc2bdf876c2ea1503/ql/src/java/org/apache/hadoop/hive/ql/exec/Utilities.java#L3369

 

This estimate is not optimal in all cases. In the worst case, it case result in a single reducer being launched , which can lead to a significant bottleneck in performance .

 

Ideally,  the number of reducers launched should equal to number of buckets, which is the case for explicitly bucketed tables.

 

 

Issue Links

• relates to: HIVE-25611 OOM when running MERGE query on wide transactional table with many buckets
• links to: GitHub Pull Request #3224