What is the difference between mlrun.code_to_function and mlrun.new_project?

What is the difference between mlrun.code_to_function and mlrun.new_project? For example, we can deploy a function with code_to_function:

a_fn = mlrun.code_to_function(name='my_function', 
                              handler='handler', 
                              kind='nuclio', 
                              image='mlrun/mlrun')
a_fn.deploy()

or we can deploy a function with mlrun.project.set_function

project = mlrun.new_project(project_name_base,
                            context=project_path,                           
                            user_project=True)                                                    
fn = project.set_function("my_function.ipynb", 
                           name='my_function')

In which case I should use code_to_function or set_function?

Solution

One of the core concepts of MLRun is creating a serverless function out of a piece of code. You can specify a Python file, entrypoint function, Docker image, K8s resources, and more. code_to_function is how this is accomplished. See this page in the docs for more info.

# file.py
def handler(context):
    context.logger.info("Hello World")

# deploy.py
from mlrun import code_to_function, new_project

project = new_project(name="my-project")

fn = code_to_function(
    name="my-function",
    project="my-project",
    filename="file.py",
    kind="job",
    handler="handler",
    image="mlrun/mlrun"
)
fn.run()

You can create and run these functions on their own, or use something like KubeFlow to orchestrate a pipeline with multiple functions. set_project is part of that workflow. You can use the function created via code_to_function or just specify some of the parameters within set_project. Then, you will be able to use this function as a part of a larger KubeFlow pipeline. See this page in the docs for more info.

project.set_function(fn)

project.set_function(
    name="my-function",
    func="file.py",
    kind="job",
    image="mlrun/mlrun"
)