[IMPALA-9176] Make access to null-aware partition from PartitionedHashJoinNode read-only

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: Impala 4.0.0
• Component/s: Backend
• Labels: multithreading
• Epic Link: Impala Multi-threaded query execution
• Epic Color: ghx-label-7
• Assignee: Tim Armstrong
• Reporter: Tim Armstrong
• Created: 20/Nov/19 16:52
• Updated: 27/Apr/20 23:58
• Resolved: 24/Apr/20 22:48

Description

Currently the accesses to null_aware_partition() are logically read-only (since the rows and other state is not mutated) and only accesses the build row when pinned, but is implemented using the built-in read iterator of BufferedTupleStream. This would prevent sharing of the build side for null-aware anti-join.

We need to either allow multiple read iterators for a pinned stream, or build an auxiliary structure, e.g. an array of Tuple ptrs or FlatRowPtr.

Issue Links

• breaks: IMPALA-9701 data race detected in ConcurrentReaders test in TSAN build
• is depended upon by: IMPALA-9156 Share broadcast join builds between fragments