[IMPALA-5586] Null-aware anti-join can take a long time to cancel

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: Impala 2.10.0
• Fix Version/s: Impala 2.10.0
• Component/s: Backend
• Labels: ramp-up
• Target Version: Product Backlog
• Epic Color: ghx-label-3
• Assignee: Anuj Phadke
• Reporter: Tim Armstrong
• Created: 27/Jun/17 17:44
• Updated: 26/Jul/17 16:35
• Resolved: 26/Jul/17 16:35

Description

Queries with a null-aware anti-join joining on a large number of NULLs can take a long time to cancel if threads are stuck in PartitionedHashJoinNode::EvaluateNullProbe().

Reproduction:

1. Run this query against tpch_parquet:

   select l_orderkey, l_partkey, l_suppkey, l_linenumber
   from lineitem
   where if(l_orderkey % 2 = 0, NULL, l_partkey) not in (
     select p_partkey
     from part
     where p_retailprice < l_extendedprice * l_tax)
   order by 1,2,3,4
   

2. Cancel it
3. See Impala still consuming CPU in EvaluateNullProbe() for a long time after cancellation