TypeDescriptor information gets lost when applying multiple DoFn on Composite Transform

[damccorm]

While working on BEAM-12384 to preserve correctly the typeDescriptor of the Read transform. I noticed that ParDo passes the outputType information into the next transforms by taking the output of getOutputTypeDescriptor specified in the DoFn, but the subsequent transform does not get the full inputTypeDescriptor.

There are other open questions:

1. How can we make this easier for users who rarely override the getOutputTypeDescriptor?
2. Where the typeDescriptor should come in the SDF case, from the Restriction?

For more details:
#14854 (comment)

 

 

 

Imported from Jira BEAM-12420. Original Jira may contain additional context.
Reported by: iemejia.