When multiple nodes concurrently return large tasks to the driver, an OutOfMemoryError can be raised. I believe this happens because in this scenario the call to the method that determines whether a serialized object read from the network connection will fit in memory is separate from the actual reading of the object, i.e. they are not performed atomically within a "single transaction".

What I think is happening is that while a thread is reading a serialized task from a node, another thread evaluates that another tasks will fit in memory, which will be false once the task from the first thread has been entirely stored in memory. Hence the second thread reads another task and stores it in memory, instead of using the disk overflow, which triggers the OOME.