Zhanghao Chen created FLINK-37680:
-------------------------------------
Summary: Add subclass cache for PojoSerializer
Key: FLINK-37680
URL: https://issues.apache.org/jira/browse/FLINK-37680
Project: Flink
Issue Type: Sub-task
Components: API / Type Serialization System
Reporter: Zhanghao Chen
Attachments: image-2025-04-15-16-31-55-289.png,
image-2025-04-15-16-33-11-325.png
When the value to deserialize is an untagged subclass of the declared POJO
type, Flink loads its acutal class every time using {{{}Class.forName{}}},
which is inefficient. We can cache the subclasses to speed up.
We tested on a production job, and the result is promising:
* Flame graph before the optim (Class.forName highlighted in green):
!image-2025-04-15-16-31-55-289.png|width=457,height=150!
* Flame graph after the optim (Subclass cache access highlighted in green):
!image-2025-04-15-16-33-11-325.png|width=460,height=156!
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
