Can you leverage the environments similar to diabetes_regression_build_train_pipeline.py

Done

Why is this AKS config deleted ?

Environment definition is moved to Python script.
AML Python SDK does not provide a way to define an environment in a YAML file, only in code (I have checked - the YAML parsing functionality is implemented directly in the AZ CLI ML extension - would be a good thing to propose to add in the SDK!)

Please change this to CLI and config files instead of SDK to deploy model. We want to demonstrate the ease of CLI to deploy models

We would still not have YML files anymore, and need an SDK script to generate the environment, and put the name in a json. Then we would be able to call the CLI. We would essentially remove a line from the existing script, but then have a delay while the extension gets installed every time for every build. Having a program generate input for a CLI would not seem to be a meaningful option to users.

Ideally the service itself should support such declarative environment versioning, rather than the imperative (additive) version behavior available today which is hard to use with DevOps.

where is this file used?

In Environment.load_from_directory(). Added comment to script for clarity fe35f83

azureml_environment.json is not pointing to any conda specification file

That is the file as generated by az ml environment scaffold.

I assume your concern is about the null values in this section of the file:

    "python": {
        "userManagedDependencies": false,
        "interpreterPath": "python",
        "condaDependenciesFile": null,
        "baseCondaEnvironment": null
    },

If you look at the AML SDK source code, load_from_directory() actually fills the conda_dependencies from a file named conda_dependencies.yml is one exists in the passed directory:

    def load_from_directory(path):
        """Load an environment definition from the files in a directory.

        :param path: Path to the source directory.
        :type path: str
        """
        definition_path = os.path.join(path, _DEFINITION_FILE_NAME)
        if not os.path.isfile(definition_path):
            raise FileNotFoundError(definition_path)

        with open(definition_path, "r") as definition:
            environment_dict = json.load(definition)
        env = Environment._deserialize_and_add_to_object(environment_dict)

        conda_file_path = os.path.join(path, _CONDA_DEPENDENCIES_FILE_NAME)
        if os.path.isfile(conda_file_path):
            env.python = env.python or PythonSection(_skip_defaults=True)
            env.python.conda_dependencies = CondaDependencies(conda_file_path)

        base_dockerfile_path = os.path.join(path, _BASE_DOCKERFILE_FILE_NAME)
        if os.path.isfile(base_dockerfile_path):
            with open(base_dockerfile_path, "r") as base_dockerfile:
                env.docker = env.docker or DockerSection(_skip_defaults=True)
                env.docker.base_dockerfile = base_dockerfile.read()

I've tested that things works as expected, as when I uncomment r-essentials from the conda file and run ml_service/pipelines/diabetes_regression_build_train_pipeline_with_r.py, training succeeds as shown by R output in the logs.