[SPARK-18507] Major performance regression in SHOW PARTITIONS on partitioned Hive tables

• Type: Sub-task
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 2.1.0
• Fix Version/s: 2.1.0
• Component/s: SQL
• Target Version/s: 2.1.0
• Assignee: Wenchen Fan
• Reporter: Michael MacFadden
• Created: 19/Nov/16 00:27
• Updated: 29/Nov/16 21:27
• Resolved: 22/Nov/16 20:26

Description

Commit ccb11543048dccd4cc590a8db1df1d9d5847d112 (https://github.com/apache/spark/commit/ccb11543048dccd4cc590a8db1df1d9d5847d112) appears to have introduced a major regression in the performance of the Hive SHOW PARTITIONS command. Running that command on a Hive table with 17,337 partitions in the spark-sql shell with the parent commit of ccb1154 takes approximately 7.3 seconds. Running the same command with commit ccb1154 takes approximately 250 seconds.

I have not had the opportunity to complete a thorough investigation, but I suspect the problem lies in the diff hunk beginning at https://github.com/apache/spark/commit/ccb11543048dccd4cc590a8db1df1d9d5847d112#diff-159191585e10542f013cb3a714f26075L675. If that's the case, this performance issue should manifest itself in other areas as this programming pattern was used elsewhere in this commit.

Issue Links

• is related to: SPARK-18545 Verify number of hive client RPCs in PartitionedTablePerfStatsSuite
• links to: [Github] Pull Request #15978 (cloud-fan)