[
https://issues.apache.org/jira/browse/FLINK-19052?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Piotr Nowojski updated FLINK-19052:
-----------------------------------
Priority: Not a Priority (was: Minor)
> Performance issue with PojoSerializer
> -------------------------------------
>
> Key: FLINK-19052
> URL: https://issues.apache.org/jira/browse/FLINK-19052
> Project: Flink
> Issue Type: Improvement
> Components: API / Type Serialization System
> Affects Versions: 1.11.1
> Environment: Flink 1.12 master on 26.08.2020
> Reporter: Roman Grebennikov
> Priority: Not a Priority
> Labels: auto-deprioritized-major, pull-request-available,
> stale-minor
> Attachments: image-2020-08-26-10-46-19-800.png,
> image-2020-08-26-10-49-59-400.png
>
>
> Currently PojoSerializer.createInstance() uses reflection call to create a
> class instance. As this method is called on each stream element on
> deserialization, reflection overhead can become noticeable in
> serialization-bound cases when:
> # Pojo class is small, so instantiation is noticeable.
> # The job is not having heavy CPU-bound event processing.
> See this flamegraph built for
> flink-benchmarks/SerializationFrameworkMiniBenchmarks.serializerPojo
> benchmark:
> !image-2020-08-26-10-46-19-800.png!
> This Reflection.getCallerClass method consumes a lot of CPU, mostly doing a
> security check if we allowed to do this reflective call.
>
> There is no true reason to perform this check on each deserializing event, so
> to speed things up we can just cache the constructor using MetaHandle, so
> this check will be performed only once. With this tiny fix, the
> getCallerClass is gone:
>
> !image-2020-08-26-10-49-59-400.png!
>
> The benchmark result:
> {noformat}
> serializerPojo thrpt 100 487.706 ± 30.480 ops/ms
> serializerPojo thrpt 100 569.828 ± 8.815 ops/m{noformat}
> Which is +15% to throughput.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
