Added dev. container

[stemann]

Contributes to #31 - notably by changing default URI to be ENV["MLFLOW_URI"] instead of http://localhost:5000

[stemann]

Updated wrt. v0.4.6

[ablaom]

@stemann Thanks for this contribution.

I don't understand the .devcontainer business. I guess this is Windows-specific?

However, the other changes indeed address #31 from my point-of-view (have tested this).

@pebeto @deyandyankov Can one of you comment on the addition of the .devcontainer folder?

@stemann If we don't get a response in the next week or so, please ping.

[pebeto]

Hi.

By the addition of .devcontainer, you are configuring a Github Codespace. Please expand about the usage, because we need to understand your modifications.

The MLFLOW_URI doesn't need to be applied as a default value in MLFlow construct. It was just applied to the test.

[pebeto]

I don't consider this as a way to go. The default location must be the mlflow default location (http://localhost:5000).

MLFlowClient.jl is just a communication layer, directly related to mlflow and not to MLJ. Everything related to MLJ must be applied in MLJFlow.jl. However, the way of adding ENV["MLFLOW_URI"] in terms of testing is correct, and discussed and implemented in JuliaAI/MLJFlow.jl#30.

[stemann]

The default URI is only http://localhost:5000 if running an MLFlow instance on localhost:5000 - this is not the case if running in a docker-compose set-up, and not the case in a GitLab CI context - just to mention a few examples.

But this part of the change was indeed a bit half-baked. Querying the env. var without checking if the env. var. is set is definitely bad. But I would argue that the implementation should use an env. var. if one such is set.

The Python API seems to query the env. var MLFLOW_TRACKING_URI if no tracking_uri is provided:

• https://github.com/mlflow/mlflow/blob/v2.11.3/mlflow/tracking/client.py#L92
• https://github.com/mlflow/mlflow/blob/v2.11.3/mlflow/tracking/_tracking_service/utils.py#L88-L117
• https://github.com/mlflow/mlflow/blob/v2.11.3/mlflow/environment_variables.py#L81
• https://mlflow.org/docs/latest/getting-started/logging-first-model/step2-mlflow-client.html

[pebeto]

Thanks for the info. We need to ensure that the end-user will get a fully working environment as default. That's why we define mlflow default URI and port. If the end-user changes those values or uses GitLab CI endpoints, that's something the end-user need to be aware of. That's not our business.

Please don't use MLFLOW_URI and rename it into MLFLOW_TRACKING_URI to match names. Also, as you mention, a check is needed. It can be something like:

MLFlow() = MLFlow(
    haskey(ENV, "MLFLOW_TRACKING_URI") ? ENV["MLFLOW_TRACKING_URI"] : "http://localhost:5000",
    2.0, Dict())

[stemann]

Done: Compare 5fac59d to 56479a5

[stemann]

The dev. container was primarily intended for running locally (on any platform) - it enables VS Code (and other editors?) to open a dev. env. - in this case with a running MLFlow instance - ready for running tests.

[stemann]

@pebeto Awaiting approval for tests to run.

[pebeto]

lgtm