The current code path for accessing Spark DataFrame data in Python using PySpark passes through an inefficient serialization-deserialiation process that I've examined at a high level here: https://gist.github.com/wesm/0cb5531b1c2e346a0007. Currently, RDD[Row] objects are being deserialized in pure Python as a list of tuples, which are then converted to pandas.DataFrame using its from_records alternate constructor. This also uses a large amount of memory.
For flat (no nested types) schemas, the Apache Arrow memory layout (https://github.com/apache/arrow/tree/master/format) can be deserialized to pandas.DataFrame objects with comparatively small overhead compared with memcpy / system memory bandwidth – Arrow's bitmasks must be examined, replacing the corresponding null values with pandas's sentinel values (None or NaN as appropriate).
I will be contributing patches to Arrow in the coming weeks for converting between Arrow and pandas in the general case, so if Spark can send Arrow memory to PySpark, we will hopefully be able to increase the Python data access throughput by an order of magnitude or more. I propose to add an new serializer for Spark DataFrame and a new method that can be invoked from PySpark to request a Arrow memory-layout byte stream, prefixed by a data header indicating array buffer offsets and sizes.