cirq_google gets incompatible protobuf version

[95-martin-orion]

Description of the issue

cirq_google prerelease is broken - see error message below

How to reproduce the issue

Run the following code in a fresh Colab:

!pip install -q cirq-google --pre
import cirq_google

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
tensorflow 2.8.0 requires tf-estimator-nightly==2.8.0.dev2021122109, which is not installed.
tensorflow-metadata 1.7.0 requires protobuf<4,>=3.13, but you have protobuf 4.21.0rc1 which is incompatible.
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
[<ipython-input-2-6fbbeec494be>](https://localhost:8080/#) in <module>()
      1 get_ipython().system('pip install -q cirq-google --pre')
----> 2 import cirq_google

5 frames
[/usr/local/lib/python3.7/dist-packages/cirq_google/__init__.py](https://localhost:8080/#) in <module>()
     17 import sys
     18 from cirq import _compat
---> 19 from cirq_google import api
     20 
     21 from cirq_google._version import __version__

[/usr/local/lib/python3.7/dist-packages/cirq_google/api/__init__.py](https://localhost:8080/#) in <module>()
     14 """Helpers for converting to/from api data formats."""
     15 
---> 16 from cirq_google.api import v2

[/usr/local/lib/python3.7/dist-packages/cirq_google/api/v2/__init__.py](https://localhost:8080/#) in <module>()
     14 """Data format v2 for google api."""
     15 
---> 16 from cirq_google.api.v2 import (
     17     batch_pb2,
     18     calibration_pb2,

[/usr/local/lib/python3.7/dist-packages/cirq_google/api/v2/batch_pb2.py](https://localhost:8080/#) in <module>()
     12 
     13 
---> 14 from . import program_pb2 as cirq__google_dot_api_dot_v2_dot_program__pb2
     15 from . import result_pb2 as cirq__google_dot_api_dot_v2_dot_result__pb2
     16 from . import run_context_pb2 as cirq__google_dot_api_dot_v2_dot_run__context__pb2

[/usr/local/lib/python3.7/dist-packages/cirq_google/api/v2/program_pb2.py](https://localhost:8080/#) in <module>()
     36       serialized_options=None,
     37       type=None,
---> 38       create_key=_descriptor._internal_create_key),
     39     _descriptor.EnumValueDescriptor(
     40       name='MOMENT_BY_MOMENT', index=1, number=1,

[/usr/local/lib/python3.7/dist-packages/google/protobuf/descriptor.py](https://localhost:8080/#) in __new__(cls, name, index, number, type, options, serialized_options, create_key)
    753                 type=None,  # pylint: disable=redefined-builtin
    754                 options=None, serialized_options=None, create_key=None):
--> 755       _message.Message._CheckCalledFromGeneratedFile()
    756       # There is no way we can build a complete EnumValueDescriptor with the
    757       # given parameters (the name of the Enum is not known, for example).

AttributeError: module 'google._upb._message' has no attribute 'Message'

Cirq version
0.15.0.dev20220512003129

[95-martin-orion]

As the error message suggests, getting a matching protobuf version is a valid workaround. If you are experiencing this issue, adding !pip install protobuf~=3.13 before the cirq_google installation line resolves the issue.

[maffoo]

4.21.0rc1 is a pre-release; the latest protobuf release on pypi is 3.20.1: https://pypi.org/project/protobuf/#history. I guess passing --pre to pip causes it to fetch prereleases for all deps, not just for cirq-google itself. We could change the protobuf dependency on cirq-google to add a <4 clause like tensorflow has.

[dabacon]

This appears to now be blocking our CI. For reasons I don't quite understand yet, our isolated notebook tests have not been running during CI, but now appear to be and failing. See #5383

[dabacon]

Ended up that a notebook test with --pre was what was causing this. Definitely need to fix before next release.