[
https://issues.apache.org/jira/browse/SPARK-25258?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16595650#comment-16595650
]
Sean Owen commented on SPARK-25258:
-----------------------------------
Both of these suggest kryo 4.
The issue is, does it cause any breaking changes. I don't so far know of any
problems, but that's the kind of input that would be great to confirm a change
like this, from those that may be watching Kryo closely.
> Upgrade kryo package to version 4.0.2+
> --------------------------------------
>
> Key: SPARK-25258
> URL: https://issues.apache.org/jira/browse/SPARK-25258
> Project: Spark
> Issue Type: Wish
> Components: Spark Core
> Affects Versions: 2.1.0, 2.3.1
> Reporter: liupengcheng
> Priority: Major
>
> Recently, we encountered a kryo performance issue in spark2.1.0, and the
> issue affect all kryo below 4.0.2, so it seems that all spark version might
> encounter this issue.
> Issue description:
> In shuffle write phase or some spilling operation, spark will use kryo
> serializer to serialize data if `spark.serializer` is set to
> `KryoSerializer`, however, when data contains some extremely large records,
> kryoSerializer's MapReferenceResolver would be expand, and it's `reset`
> method will take a long time to reset all items in writtenObjects table to
> null.
> com.esotericsoftware.kryo.util.MapReferenceResolver
> {code:java}
> public void reset () {
> readObjects.clear();
> writtenObjects.clear();
> }
> public void clear () {
> K[] keyTable = this.keyTable;
> for (int i = capacity + stashSize; i-- > 0;)
> keyTable[i] = null;
> size = 0;
> stashSize = 0;
> }
> {code}
> I checked the kryo project in github, and this issue seems fixed in 4.0.2+
> [https://github.com/EsotericSoftware/kryo/commit/77935c696ee4976963aa5c6ac53d53d9b40b8bdd#diff-215fa9846e1e4e54bbeede0500de1e28]
>
> I was wondering if we can make spark kryo package upgrade to 4.0.2+ to fix
> this problem.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
