[IMPALA-5965] Avoid per-value switch on NeedsConversionInline() when decoding dictionary-encoded strings and timestamps

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: Impala 2.10.0
• Fix Version/s: Impala 2.11.0
• Component/s: Backend
• Labels: perf
• Target Version: Product Backlog
• Epic Color: ghx-label-9
• Assignee: Tim Armstrong
• Reporter: Tim Armstrong
• Created: 20/Sep/17 20:44
• Updated: 22/Sep/17 15:27
• Resolved: 22/Sep/17 15:27

Description

I noticed a minor inefficiency in the handling of NeedsConversion() in the parquet scanner. In cases where it's not a runtime constant like dictionary-encoded strings and timestamps, we switch per value. This is probably only a few instructions but in this part of the code that matters.

I did a quick benchmark and saw speedups from ~2.25s->2.11s in scan time on this query:

use tpch_parquet; 
set num_nodes=1;
set mt_dop=1;
select min(l_returnflag), min(l_linestatus) from biglineitem;
summary;