Support passing artifact_uri in config

[Calychas]

Description

Allow passing in artifact_uri to enable automatic creation of experiments with it and not defaulting to the one provided by the MLFlow Server.

Context

If the MLFlow Server has artifact_uri configured, but e.g.:

• points to the "wrong" bucket (misconfigured or we don't have permissions);
• the convention of saving artifacts in the bucket is gs://<bucket_name>/<experiment_name>/artifacts - subfolders per experiment;
• or you just want to save it somewhere else,

then it is impossible to have the experiment created automatically with the custom artifact_uri.

Slack thread

Possible Implementation

Currently there are 2 things missing in kedro-mlflow that need to happen:

• enable passing mlflow_artifact_uri in mlflow.yaml (next tomlflow_registry_uri and mlflow_tracking_uri)
• kedro-mlflow now uses mlflow.set_experiment (here) function that handles experiment creation, but unfortunately it doesn't support specifying artifact_uri. The switch to mlflow.create_experiment would need to happen to enable it.

Possible Alternatives

Don't implement it at all and perform the experiment creation manually with custom artifact_uri either:

• through the UI:
  
• or through the API. This approach may be possible to automate by creating a hook that runs and creates the experiment before the kedro-mlflow hook. It would also need to read the mlflow.yml

[lvijnck]

+1