Github user JoshRosen commented on a diff in the pull request:
https://github.com/apache/spark/pull/5924#discussion_r29718641
--- Diff:
core/src/main/scala/org/apache/spark/serializer/KryoSerializer.scala ---
@@ -125,6 +125,13 @@ class KryoSerializer(conf: SparkConf)
override def newInstance(): SerializerInstance = {
new KryoSerializerInstance(this)
}
+
+ private[spark] override lazy val supportsRelocationOfSerializedObjects:
Boolean = {
+ // If auto-flush is disabled, then Kryo may store references to
duplicate occurrences of objects
+ // in the stream rather than writing those objects' serialized bytes,
breaking relocation. See
+ // https://groups.google.com/d/msg/kryo-users/6ZUSyfjjtdo/FhGG1KHDXPgJ
for more details.
+ newInstance().asInstanceOf[KryoSerializerInstance].getAutoReset()
--- End diff --
I couldn't seem to find a good way to determine whether auto-reset was
enabled short of actually creating a serializer instance. Fortunately, it's
fairly cheap to create a KryoSerializerInstance that you don't write to because
its buffers are allocated lazily. Combined with #5606, which enables serializer
re-use in many circumstances, I don't think that this will carry a huge
performance penalty (especially since this is a `lazy val`).
I suppose that one alternative would be to move this method from
`Serializer` to `SerializerInstance`. I chose not to take that approach because
my other patch, #5868, checks serializer properties inside of
`ShuffleManager.getWriter` in order to choose which writer implementation to
use; at that call site, we don't have a SerializerInstance, so I'd still have
to create an unnecessary serializer instance there.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
