[IMPALA-1983] Impala should handle corrupted table stats better

• Type: Improvement
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 2.1, Impala 2.2
• Fix Version/s: Impala 2.2.7, Impala 2.3.0
• Labels: supportability usability
• Target Version: Impala 2.2.7, Impala 2.3.0
• Assignee: Martin Grund
• Reporter: Juan Yu
• Created: 07/May/15 19:08
• Updated: 20/Dec/15 00:05
• Resolved: 26/Aug/15 22:23

Description

We see table stats get messed up from time to time, not sure if it's due to upgrading or anything else.
The table has data but the cardinality becomes 0. This could cause Impala to choose wrong join order and affect query performance.

A warning to inform the user the table stats could be corrupted will be helpful.
Also if possible Impala should check if the cardinality is valid instead of simply choosing single node plan when cardinality == 0.

Issue Links

• relates to: IMPALA-1979 Bad Planning: Scanning TBs of data with only ONE thread