Ramin Gharib created FLINK-41241:
------------------------------------
Summary: Remember the field index in VARIANT key lookups
Key: FLINK-41241
URL: https://issues.apache.org/jira/browse/FLINK-41241
Project: Flink
Issue Type: Improvement
Components: Table SQL / Planner, API / Core
Reporter: Ramin Gharib
Fix For: 2.5.0
VARIANT field access by a literal key, and the cast from VARIANT to ROW, look
up each object field by name on every row: encode the key to UTF-8, then search
the field list. Objects of one stream mostly share their keys, so the field is
usually at the same index as in the previous row.
{code:sql}
SELECT v['user'], v['ts'], v['kind'], v['z'] FROM events
{code}
Add an internal VariantFieldGetter that encodes its key once and checks the
index where it found the key in the previous row before it searches. The index
is a hint that is always verified, so a getter can be shared between threads.
The generated code keeps one getter per field access with a literal key, and
the cast to ROW and STRUCTURED keeps one per field.
JMH, JDK 17. Per row means the 4 lookups of the query above:
||Stream of objects||Today, per row||With getters, per row||
|Same 13 keys in every row|214 ns|44 ns|
|Same 40 keys in every row|158 ns|44 ns|
|14 to 20 keys, so the looked-up key moves in most rows|296 ns|242 ns|
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
