[SPARK-3649] Remove GraphX custom serializers #2503
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
QA tests have started for PR 2503 at commit
|
|
QA tests have finished for PR 2503 at commit
|
|
Test FAILed. |
|
retest this please |
|
QA tests have started for PR 2503 at commit
|
|
QA tests have finished for PR 2503 at commit
|
|
Test PASSed. |
|
merging this. thanks! |
asfgit
pushed a commit
that referenced
this pull request
Nov 11, 2014
As [reported][1] on the mailing list, GraphX throws
```
java.lang.ClassCastException: java.lang.Long cannot be cast to scala.Tuple2
at org.apache.spark.graphx.impl.RoutingTableMessageSerializer$$anon$1$$anon$2.writeObject(Serializers.scala:39)
at org.apache.spark.storage.DiskBlockObjectWriter.write(BlockObjectWriter.scala:195)
at org.apache.spark.util.collection.ExternalSorter.spillToMergeableFile(ExternalSorter.scala:329)
```
when sort-based shuffle attempts to spill to disk. This is because GraphX defines custom serializers for shuffling pair RDDs that assume Spark will always serialize the entire pair object rather than breaking it up into its components. However, the spill code path in sort-based shuffle [violates this assumption][2].
GraphX uses the custom serializers to compress vertex ID keys using variable-length integer encoding. However, since the serializer can no longer rely on the key and value being serialized and deserialized together, performing such encoding would either require writing a tag byte (costly) or maintaining state in the serializer and assuming that serialization calls will alternate between key and value (fragile).
Instead, this PR simply removes the custom serializers. This causes a **10% slowdown** (494 s to 543 s) and **16% increase in per-iteration communication** (2176 MB to 2518 MB) for PageRank (averages across 3 trials, 10 iterations per trial, uk-2007-05 graph, 16 r3.2xlarge nodes).
[1]: http://apache-spark-user-list.1001560.n3.nabble.com/java-lang-ClassCastException-java-lang-Long-cannot-be-cast-to-scala-Tuple2-td13926.html#a14501
[2]: https://github.com/apache/spark/blob/f9d6220c792b779be385f3022d146911a22c2130/core/src/main/scala/org/apache/spark/util/collection/ExternalSorter.scala#L329
Author: Ankur Dave <ankurdave@gmail.com>
Closes #2503 from ankurdave/SPARK-3649 and squashes the following commits:
a49c2ad [Ankur Dave] [SPARK-3649] Remove GraphX custom serializers
(cherry picked from commit 300887b)
Signed-off-by: Reynold Xin <rxin@databricks.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
As reported on the mailing list, GraphX throws
when sort-based shuffle attempts to spill to disk. This is because GraphX defines custom serializers for shuffling pair RDDs that assume Spark will always serialize the entire pair object rather than breaking it up into its components. However, the spill code path in sort-based shuffle violates this assumption.
GraphX uses the custom serializers to compress vertex ID keys using variable-length integer encoding. However, since the serializer can no longer rely on the key and value being serialized and deserialized together, performing such encoding would either require writing a tag byte (costly) or maintaining state in the serializer and assuming that serialization calls will alternate between key and value (fragile).
Instead, this PR simply removes the custom serializers. This causes a 10% slowdown (494 s to 543 s) and 16% increase in per-iteration communication (2176 MB to 2518 MB) for PageRank (averages across 3 trials, 10 iterations per trial, uk-2007-05 graph, 16 r3.2xlarge nodes).