[SPARK-22073] Add workaround for HIVE-15653 to avoid stats being accidentally wiped

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 2.2.0
• Component/s: SQL
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Alexander Behm
• Created: 20/Sep/17 05:22
• Updated: 21/May/19 04:14
• Resolved: 21/May/19 04:14

Description

The alterTable() RPC to the Hive Metastore has the subtle and undesirable side-effect of sometimes wiping/recreating useful stats like 'numRows' and 'totalSize'.

Further, the alterTable() causes extra "hidden" load on the NameNode/S3/whatever to list all the files and get the 'totalSize' and 'numFiles'.

This behavior can lead to confusion, e.g., if a SparkSQL query suddenly becomes slow due to a bad plan because somebody ran an unrelated ALTER TABLE from Hive/SparkSQL which wiped the stats (not fun to do the RCA).

One way to avoid this problem is to always add a special table property DO_NOT_UPDATE_STATS to the altered table before doing the alterTable() RPC to the Metastore.

See HIVE-15653 for more details.