[IMPALA-3046] Allow per-row filters to evaluate non-SlotRef exprs

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: Impala 2.5.0
• Fix Version/s: Impala 2.5.0
• Component/s: Distributed Exec
• Labels: runtime-filters
• Target Version: Impala 2.5.0
• Assignee: Henry Robinson
• Reporter: Henry Robinson
• Created: 21/Feb/16 06:07
• Updated: 25/Feb/16 14:57
• Resolved: 25/Feb/16 14:57

Description

Per-row filters currently can not evaluate non-slotref expressions, but the frontend generates them. In some cases, the parquet scanner does not correctly ignore those filters.

The best fix is to evaluate row filters once a row has been materialized. This concentrates filter processing per row (better cache locality), and ensures that a) all slots are available for filter expr evaluation and b) there is a materialized Tuple available which means all exprs can be evaluated.

Thankfully the changes required to do this are very simple - just move logic from the column reader to HdfsParquetScanner::ReadRow().