[
https://issues.apache.org/jira/browse/SPARK-60330?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ruifeng Zheng resolved SPARK-60330.
-----------------------------------
Fix Version/s: 4.4.0
Resolution: Fixed
Issue resolved by pull request 59539
[https://github.com/apache/spark/pull/59539]
> Register classification tree nodes with Kryo
> --------------------------------------------
>
> Key: SPARK-60330
> URL: https://issues.apache.org/jira/browse/SPARK-60330
> Project: Spark
> Issue Type: Improvement
> Components: ML, Spark Core
> Affects Versions: 5.0.0, 4.4.0
> Reporter: Luka Erceg
> Assignee: Luka Erceg
> Priority: Minor
> Labels: pull-request-available
> Fix For: 4.4.0
>
>
> Spark's built-in Kryo registrations omit the classes needed to serialize
> classification tree node graphs. With spark.kryo.registrationRequired=true,
> serializing an Array[org.apache.spark.ml.tree.Node] fails unless users
> register these built-in classes manually. Without strict registration, Kryo
> writes class-name metadata for the unregistered types.
> Register Node arrays, InternalNode, LeafNode, ContinuousSplit,
> CategoricalSplit, GiniCalculator, and EntropyCalculator using the existing
> optional-class loading mechanism, preserving the separation between Spark
> Core and MLlib.
> The PR adds NodeKryoSuite to verify round-trip serialization with strict
> registration enabled for Gini and entropy trees, checking predictions, leaf
> indices, and class-count statistics across continuous and categorical splits.
> PR: https://github.com/apache/spark/pull/59539
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
