Hi Spark devs, I’ve posted an issue on JIRA ( https://issues.apache.org/jira/browse/SPARK-2878) which occurs when using Kryo serialisation with a custom Kryo registrator to register custom classes with Kryo. This is an insidious issue that non-deterministically causes Kryo to have different ID number => class name maps on different nodes, which then causes weird exceptions (ClassCastException, ClassNotFoundException, ArrayIndexOutOfBoundsException) at deserialisation time. I’ve created a reliable reproduction for the issue here: https://github.com/GrahamDennis/spark-kryo-serialisation
I’m happy to try and put a pull request together to try and address this, but it’s not obvious to me the right way to solve this and I’d like to get feedback / ideas on how to address this. The root cause of the problem is a "Failed to run spark.kryo.registrator” error which non-deterministically occurs in some executor processes during operation. My custom Kryo registrator is in the application jar, and it is accessible on the worker nodes. This is demonstrated by the fact that most of the time the custom kryo registrator is successfully run. What’s happening is that Kryo serialisation/deserialisation is happening most of the time on an “Executor task launch worker” thread, which has the thread's class loader set to contain the application jar. This happens in `org.apache.spark.executor.Executor.TaskRunner.run`, and from what I can tell, it is only these threads that have access to the application jar (that contains the custom Kryo registrator). However, the ConnectionManager threads sometimes need to serialise/deserialise objects to satisfy “getBlock” requests when the objects haven’t previously been serialised. As the ConnectionManager threads don’t have the application jar available from their class loader, when it tries to look up the custom Kryo registrator, this fails. Spark then swallows this exception, which results in a different ID number —> class mapping for this kryo instance, and this then causes deserialisation errors later on a different node. A related issue to the issue reported in SPARK-2878 is that Spark probably shouldn’t swallow the ClassNotFound exception for custom Kryo registrators. The user has explicitly specified this class, and if it deterministically can’t be found, then it may cause problems at serialisation / deserialisation time. If only sometimes it can’t be found (as in this case), then it leads to a data corruption issue later on. Either way, we’re better off dying due to the ClassNotFound exception earlier, than the weirder errors later on. I have some ideas on potential solutions to this issue, but I’m keen for experienced eyes to critique these approaches: 1. The simplest approach to fixing this would be to just make the application jar available to the connection manager threads, but I’m guessing it’s a design decision to isolate the application jar to just the executor task runner threads. Also, I don’t know if there are any other threads that might be interacting with kryo serialisation / deserialisation. 2. Before looking up the custom Kryo registrator, change the thread’s class loader to include the application jar, then restore the class loader after the kryo registrator has been run. I don’t know if this would have any other side-effects. 3. Always serialise / deserialise on the existing TaskRunner threads, rather than delaying serialisation until later, when it can be done only if needed. This approach would probably have negative performance consequences. 4. Create a new dedicated thread pool for lazy serialisation / deserialisation that has the application jar on the class path. Serialisation / deserialisation would be the only thing these threads do, and this would minimise conflicts / interactions between the application jar and other jars. #4 sounds like the best approach to me, but I think would require considerable knowledge of Spark internals, which is beyond me at present. Does anyone have any better (and ideally simpler) ideas? Cheers, Graham
