TheNeuralBit commented on a change in pull request #12882:
URL: https://github.com/apache/beam/pull/12882#discussion_r492219317
##########
File path: sdks/python/apache_beam/dataframe/schemas.py
##########
@@ -55,17 +159,149 @@ def expand(self, pcoll):
lambda batch: pd.DataFrame.from_records(batch, columns=columns))
-def _make_empty_series(name, typ):
- try:
- return pd.Series(name=name, dtype=typ)
- except TypeError:
- raise TypeError("Unable to convert type '%s' for field '%s'" % (name, typ))
+def _make_proxy_series(name, typehint):
+ # Default to np.object. This is lossy, we won't be able to recover the type
+ # at the output.
+ dtype = BEAM_TO_PANDAS.get(typehint, np.object)
+
+ return pd.Series(name=name, dtype=dtype)
def generate_proxy(element_type):
# type: (type) -> pd.DataFrame
- return pd.DataFrame({
- name: _make_empty_series(name, typ)
- for name,
- typ in named_fields_from_element_type(element_type)
- })
+
+ """ Generate a proxy pandas object for the given PCollection element_type.
+
+ Currently only supports generating a DataFrame proxy from a schema-aware
+ PCollection."""
+ fields = named_fields_from_element_type(element_type)
+ return pd.DataFrame(
+ {name: _make_proxy_series(name, typehint)
+ for name, typehint in fields},
+ columns=[name for name, _ in fields])
+
+
+def element_type_from_proxy(proxy):
+ # type: (pd.DataFrame) -> type
+
+ """ Generate an element_type for an element-wise PCollection from a proxy
+ pandas object. Currently only supports converting the element_type for
+ a schema-aware PCollection to a proxy DataFrame.
+
+ Currently only supports generating a DataFrame proxy from a schema-aware
+ PCollection."""
+ indices = [] if proxy.index.names == (None, ) else [
Review comment:
I thought the MultiIndex or named case was important since otherwise
we'll drop the grouped column(s) when unbatching the result of a grouped
aggregation.
It raise some tricky issues though:
- Index names are not required to be unique.
- It looks like my assumption that all MultiIndexes are named is wrong. It's
possible to create a `MultiIndex` with `names=[None, None, 'foo']`, which would
break this badly.
- Type information is not necessarily preserved in indexes. e.g. Int64Index
doesn't support nulls like Series with Int64Dtype does. if one is added it's
converted to a Float64Index with nans.
Maybe including the index shouldn't be the default until we have a better
handle on these edge cases.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
