It's more flexible to use FakeDataSource, since it allows to testing different upstream behaviors (e.g., upstream not being able to resolve the content length). PiperOrigin-RevId: 362072899
