[HIVE-2035] Use block-level merge for RCFile if merging intermediate results are needed

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.8.0
• Release Note: For tables stored as RCFile, intermediate results that have too many small files will be merged with a block-level merge that does not deserialize and re-serialized the contents of each block.
• Assignee: Franklin Hu
• Reporter: Ning Zhang
• Created: 08/Mar/11 22:27
• Updated: 24/Jul/14 22:32
• Resolved: 18/Jul/11 19:41

Description

Currently if hive.merge.mapredfiles and/or hive.merge.mapfile is set to true the intermediate data could be merged using an additional MapReduce job. This could be quite expensive if the data size is large. With HIVE-1950, merging can be done in the RCFile block level so that it bypasses the (de-)compression, (de-)serialization phases. This could improve the merge process significantly.

This JIRA should handle the case where the input table is not stored in RCFile, but the destination table is (which requires the intermediate data should be stored in the same format as the destination table).

Issue Links

• relates to: HIVE-1950 Block merge for RCFile
• relates to: HIVE-7509 Fast stripe level merging for ORC