Sagar Mittal created SPARK-57760:
------------------------------------
Summary: Optimize StatefulProcessorApiClient._serialize_to_bytes
Key: SPARK-57760
URL: https://issues.apache.org/jira/browse/SPARK-57760
Project: Spark
Issue Type: Improvement
Components: Structured Streaming
Affects Versions: 4.2.0
Reporter: Sagar Mittal
In an internal benchmark, we have found `_serialize_to_bytes` is a performance
bottleneck. [Here is a microbenchmark for this
function|https://gist.github.com/funrollloops/62bf5fae1654910b63a4539e1181db91].
I identified two optimizations to improve performance:
1. Use PickleSerializer() instead of the default CPickleSerializer
(CloudPickleSerializer). The stateful processor path does not deal with
code objects so the lighter serializer is sufficient.
2. Micro-optimize state value normalization: add a fast-path for Python
primitives, prefer `map` over generator comprehensions, and hoist the
numpy import and helper function definition to module level so they are
executed once rather than per call.
Together these changes improve transform-with-state throughput on a simple
rolling-window benchmark by around ~10%.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
