[IMPALA-6280] Invalid plan for sorted INSERT with an outer join and null checking

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 2.10.0
• Fix Version/s: Impala 2.11.0
• Component/s: Frontend
• Labels: kudu planner regression
• Epic Color: ghx-label-8
• Assignee: Thomas Tauber-Marshall
• Reporter: Alexander Behm
• Created: 05/Dec/17 22:27
• Updated: 09/Dec/17 00:00
• Resolved: 09/Dec/17 00:00

Description

Kudu DML queries may fail due to an invalid plan in the following circumstances:

• Query has an outer join that involves an inline view
• The outer-joined inline view contains NULL-checking expressions like COALESCE()/IFNULL()/CASE and/or constant expressions like literals

The following error message will appear in the logs.

Invalid plan. TupleIsNullPredicate has invalid tuple idx.

Further, due to a separate issue, the coordinating impalad may crash (IMPALA-6262).

This issue is a regression introduced in Impala 2.10 by the recent addition of partial sorting before Kudu DML. See IMPALA-5498.

Reproduction

create table ktest (a int primary key, b int) stored as kudu;

upsert into ktest
select t1.id, v.id from functional.alltypes t1
left outer join (select ifnull(id, 10) id from functional.alltypessmall) v
on t1.id = v.id;

Issue Links

• is broken by: IMPALA-5498 Support for partial sorts
• is related to: IMPALA-6262 Crash Impalad [ DataSink::Create fail which cause profile nullptr ]