If a task runs on the same slave that produced its input, then we can save the work of deserialization by caching output data. Note that it will be easy to get an estimate of how much space the cache is taking by measuring the size of the serialized copy.