Details
-
Improvement
-
Status: Open
-
P3
-
Resolution: Unresolved
-
None
-
None
-
None
Description
FileBasedCacheManager used by Python's InteractiveRunner does not preserve PCoder for elements of a PCollection being cached on disk. I suggest that the cache on-disk format to be changed to TFRecords (which are supported by Beam) and FileBasedCacheManager would store the desired PCoder for cached collections.
Currently, it is not possible to work with dynamically-generated protocol buffer messages in interactive runner mode because of pickling errors.