feat: make read_session optional in ReadRowsStream.rows()

[tswast]

Is your feature request related to a problem? Please describe.

ReadRowsStream.rows() and other methods that decode the data currently take a ReadSession as a required argument. This is so that it can use the avro/arrow schema to decode the rows.

This is a bit of a pain, especially when workers are on separate processes.

Describe the solution you'd like

#166 adds arrow_schema and avro_schema properties to ReadRowsResponse. These are populated only in the first message of a read stream, but can be cached locally to decode subsequent messages.

To avoid breaking changes, a ReadSession can still be accepted, but it should be optional. Thankfully Python supports arguments which work either as args or kwargs, so I think this does not have to be a breaking change. We need to make sure we comment to the library devs that the order matters for this argument, though.

Describe alternatives you've considered

New Reader class?

Additional context

N/A

[tswast]

Although it's not documented in the Python reference, the source proto has a schema oneof field for this behavior.

https://github.com/googleapis/googleapis/blob/fe3cc5e650b8e52537f02549fe5fb49fc386f3a9/google/cloud/bigquery/storage/v1/storage.proto#L209-L215

We should be able to use very similar logic as from_read_session

python-bigquery-storage/google/cloud/bigquery_storage_v1/reader.py

Lines 495 to 504 in 8c10713

	def from_read_session(read_session):
	schema_type = read_session._pb.WhichOneof("schema")
	if schema_type == "avro_schema":
	return _AvroStreamParser(read_session)
	elif schema_type == "arrow_schema":
	return _ArrowStreamParser(read_session)
	else:
	raise TypeError(
	"Unsupported schema type in read_session: {0}".format(schema_type)
	)

[tswast]

Behavior change: Can't raise ImportError if necessary libraries aren't installed until iteration time because without read_session, we don't know the serialization format until the first ReadRowsResponse message is parsed.

[tswast]

Behavior change: Empty stream returns dataframe / arrow table with no columns because schema is not available.

Edit: Since this is a breaking change, maybe we actually do still parse read_session for this? We can still deprecate read_session but wait to remove all parsing code until next major version?