[IMPALA-11185] Reuse orc::ColumnVectorBatch in the scanner life-cycle

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: Impala 4.1.0
• Component/s: Backend
• Epic Link: ORC Performance
• Epic Color: ghx-label-1
• Assignee: Quanlong Huang
• Reporter: Quanlong Huang
• Created: 16/Mar/22 03:02
• Updated: 19/Mar/22 01:08
• Resolved: 19/Mar/22 01:08

Description

In HdfsOrcScanner::AssembleRows(), we always re-create a orc::ColumnVectorBatch. The ideal pattern is reusing the batch and only destroyed it when the scanner is closed.

In the flame graph of TPC-H Q1 collected by drorke , the createRowBatch and destructors occupies almost half of the scanner time.