INVALID_PARAMETER_VALUE: Model registry functionality is unavailable

[andrewcstewart]

If you like the repo, please give it a ⭐

Description

While trying to use kedro-mlflow with sqlite backend, I receive the following error in the UI:

INVALID_PARAMETER_VALUE: Model registry functionality is unavailable; got unsupported URI 'file:///kedro/data/06_models' for model registry data storage. Supported URI schemes are: ['postgresql', 'mysql', 'sqlite', 'mssql']. See https://www.mlflow.org/docs/latest/tracking.html#storage for how to run an MLflow server against one of the supported backend storage locations.

mlflow.yml is the default generated configuration with the following changes:

server:
  mlflow_tracking_uri: data/06_models/ # if null, will use mlflow.get_tracking_uri() as a default
  mlflow_registry_uri: data/06_models/registry.db # if null, mlflow_tracking_uri will be used as mlflow default

Happy to submit a more thorough issue w/ reproducibility information etc, but want to first sanity check if there is an obvious mistake in my understanding of the registry configuration?

[Galileo-Galilei]

Hi @andrewcstewart, actually your configuration is almost correct: when you do not specify the dialect of your URIs, they are interpreted as local uri, and so they are converted to an absolute path prepended with file:///. This actually makes senses: if you use a relative path, mlflow cannot "guess" what type of database you are using.

You need to use an absolute path, and to preprend it with sqlite:/// dialect, so your configuration becomes:

server:
  mlflow_tracking_uri: data/06_models/ # <- ok, if it is relative it will create a mlruns folder here
  mlflow_registry_uri: sqlite:///absolute/path/to/data/06_models/registry.db # <- change this line

[andrewcstewart]

Aha, that makes sense. The thought occurred to me but I thought I pulled that directly from an example in the docs and figured the plugin's automatic handling of relative paths -> abs paths must also be autodetecting .db as sqlite.

Anyway, thanks!

[Galileo-Galilei]

I think I should rather correct the documentation than trying to autodetect sqlite URI 😅 The URI validation logic is already quite complicated due to subtle differences between linux / windows / macOS.

This would make extra assumptions (is your .db really a sqlite file? there are a lot of other possibilities! Is .db the only extension I should detect? This is likely the most common, but it is not completely standardized), and I don't think Ishould go against mlflow decisions here.

I let the issue open to remember to correct the documentation which is indeed incorrect.

[andrewcstewart]

Little bit of missing context behind my presumption there is I had come across some issue over at https://github.com/mlflow/mlflow about defaulting to sqlite:// when registry_uri contains .db (or something along those lines).