Fix attribute types for autogenerated protobuf classes

[PicardParis]

Issue

• The autogenerated protobuf classes are of the following form:

  class ProtobufClass(proto.Message):
      """...
      Attributes:
          # Primitive types
          attribute_1 (str):...
          attribute_2 (int):...
          attribute_3 (float):...
          ...
          # Compound types
          attribute_A (TYPE_A):...
          attribute_B (Sequence[TYPE_B]):...
          attribute_C (Mapping[KEY_TYPE_C, VALUE_TYPE_C]):...
      """
      attribute_1 = proto.Field(proto.STRING, number=1)
      attribute_2 = proto.Field(proto.INT32, number=2)
      attribute_3 = proto.Field(proto.FLOAT, number=3)
      ...
      attribute_A = proto.Field(proto.MESSAGE, number=4, message="TYPE_A")
      attribute_B = proto.RepeatedField(proto.MESSAGE, number=5, message="TYPE_B")
      attribute_C = proto.MapField(...)
      ...

  • ✔️ The docstrings seem perfectly correct (they match the actual types at runtime).
  • ❌ The attributes indicate protobuf types (always proto.Field proto.RepeatedField, or proto.MapField?).
• As a consequence, type checkers (e.g. pyright) can report issues such as:

  • Cannot access member "MEMBER" for type "Field"
    Member "MEMBER" is unknown
    

  • Expression of type "Field" cannot be assigned to return type "TYPE"
    "Field" is incompatible with "TYPE"
    

  • "RepeatedField" is not iterable
    "__iter__" method not defined
    

  • Argument of type "RepeatedField" cannot be assigned to parameter "__obj" of type "Sized" in function "len"
    "RepeatedField" is incompatible with protocol "Sized"
    "__len__" is not present
    

Proposal

• Adding type annotations at the attribute level (attribute_X: TYPE_X = proto.*) may consistently fix these type checker issues.
• In a nutshell, here is what it could look like:

  class ProtobufClass(proto.Message):
      """...
      Attributes:
          attribute_1 (TYPE_1):...
          attribute_2 (TYPE_2):...
          attribute_3 (Sequence[TYPE_3]):...
          attribute_4 (Mapping[KEY_TYPE_4, VALUE_TYPE_4]):...
      """
      attribute_1: TYPE_1 = proto.Field(proto.PRIMITIVE_PROTO_TYPE, number=1)
      attribute_2: TYPE_2 = proto.Field(proto.MESSAGE, number=2, message="TYPE_2")
      attribute_3: Sequence[TYPE_3] = proto.RepeatedField(proto.MESSAGE, number=3, message="TYPE_3")
      attribute_4: Mapping[KEY_TYPE_4, VALUE_TYPE_4] = proto.MapField(...)
      ...

Why is it an issue

• It's likely to generate more and more friction as static type checking becomes the norm (+ the numbers of users and client libraries keep increasing).
• Some client libraries are pretty large (e.g. google.cloud.documentai.Document) and IDE autocompletion is a must-have.
• This can lead developers to add unnecessary boilerplate (e.g. typing.TypeAlias) or casting (e.g. typing.cast or worse).
• This has already been reported: SA#73126100 & gapic-generator-python#694.

[vam-google]

Thanks @PicardParis for your request. In case you feel like you can fix it yourself, feel free to do so by opening a PR (that would be greatly appreciated given we don't have enough resources to address all open issues in timely manner). The changes are expected to be done around here:

https://github.com/googleapis/gapic-generator-python/blob/main/gapic/templates/%25namespace/%25name_%25version/%25sub/types/_message.py.j2#L17

https://github.com/googleapis/gapic-generator-python/blob/main/gapic/ads-templates/%25namespace/%25name/%25version/%25sub/types/_message.py.j2#L17

[PicardParis]

Hi @vam-google, @parthea,

• I've set up the dev environment in a VM and made a few positive tests to update/generate classes. I should be able to go on and propose a PR.
• Why are there 2 sets of templates (templates and ads-templates)? Is it enough to modify templates (for google.cloud.*) or should both be updated?
• To avoid confusion, can you assign the issue to me? (I don't have any rights for this repo)