hey all, I'm having some issues related to environments/dependencies when I deploy my Kedro pipeline to Azure ML. After the pipeline is deployed succesfully, we register the trained model by using the Azure ML Studio user interface. Then, when the model is registered we create a Batch endpoint through the Azure SDK. After connecting to the registered model in the Azure SDK, we also define an environment by a conda.yaml file:

env = Environment(
    name="setup_env",
    conda_file="conda.yaml", 
    image="mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04",
)
# Register the environment
ml_client.environments.create_or_update(env)

Then the next step is to deploy model to Batch endpoint:

deployment = BatchDeployment(
    name="deployment_name",
    description="A regression model",
    environment=ml_client.environments.get("setup_env", label="latest"),
    endpoint_name=endpoint.name,
    model=model,
    compute="batch-cluster",
    instance_count=2,
    max_concurrency_per_instance=2,
    mini_batch_size=2,
    output_action=BatchDeploymentOutputAction.APPEND_ROW,
    output_file_name="predictions.csv",
    retry_settings=BatchRetrySettings(max_retries=3, timeout=300),
    logging_level="info",
)
ml_client.batch_deployments.begin_create_or_update(deployment)

Then, after submitting a batch job to this endpoint to test if the endpoint can return a prediction value, an environment error appears.

job = ml_client.batch_endpoints.invoke(
    endpoint_name=endpoint.name, 
    input=input)

ml_client.jobs.get(job.name)

The error in Azure ML studio is in the 'prepare_image' job. This is part of the std_log: 2024-12-16T153256: ERROR: Cannot install -r /azureml-environment-setup/condaenv.lqbo0q79.requirements.txt (line 1), -r /azureml-environment-setup/condaenv.lqbo0q79.requirements.txt (line 12), -r /azureml-environment-setup/condaenv.lqbo0q79.requirements.txt (line 13), -r /azureml-environment-setup/condaenv.lqbo0q79.requirements.txt (line 6), -r /azureml-environment-setup/condaenv.lqbo0q79.requirements.txt (line 7), -r /azureml-environment-setup/condaenv.lqbo0q79.requirements.txt (line 9), azureml-dataset-runtime[fuse]==1.10.0, azureml-dataset-runtime[fuse]==1.11.0, azureml-dataset-runtime[fuse]==1.11.0.post1, azureml-dataset-runtime[fuse]==1.12.0, azureml-dataset-runtime[fuse]==1.13.0, azureml-dataset-runtime[fuse]==1.14.0, azureml-dataset-runtime[fuse]==1.15.0, azureml-dataset-runtime[fuse]==1.16.0, azureml-dataset-runtime[fuse]==1.17.0, azureml-dataset-runtime[fuse]==1.18.0, azureml-dataset-runtime[fuse]==1.19.0, azureml-dataset-runtime[fuse]==1.19.0.post1, azureml-dataset-runtime[fuse]==1.20.0, azureml-dataset-runtime[fuse]==1.21.0, azureml-dataset-runtime[fuse]==1.22.0, azureml-dataset-runtime[fuse]==1.23.0, azureml-dataset-runtime[fuse]==1.24.0, azureml-dataset-runtime[fuse]==1.25.0, azureml-dataset-runtime[fuse]==1.26.0, azureml-dataset-runtime[fuse]==1.27.0, azureml-dataset-runtime[fuse]==1.28.0, azureml-dataset-runtime[fuse]==1.29.0, azureml-dataset-runtime[fuse]==1.30.0, azureml-dataset-runtime[fuse]==1.31.0, azureml-dataset-runtime[fuse]==1.32.0, azureml-dataset-runtime[fuse]==1.33.0, azureml-dataset-runtime[fuse]==1.34.0, azureml-dataset-runtime[fuse]==1.35.0, azureml-dataset-runtime[fuse]==1.36.0, azureml-dataset-runtime[fuse]==1.37.0, azureml-dataset-runtime[fuse]==1.38.0, azureml-dataset-runtime[fuse]==1.39.0, azureml-dataset-runtime[fuse]==1.40.0, azureml-dataset-runtime[fuse]==1.41.0, azureml-dataset-runtime[fuse]==1.42.0, azureml-dataset-runtime[fuse]==1.43.0, azureml-dataset-runtime[fuse]==1.43.0.post1, azureml-dataset-runtime[fuse]==1.43.0.post2, azureml-dataset-runtime[fuse]==1.44.0, azureml-dataset-runtime[fuse]==1.45.0, azureml-dataset-runtime[fuse]==1.46.0, azureml-dataset-runtime[fuse]==1.47.0, azureml-dataset-runtime[fuse]==1.48.0, azureml-dataset-runtime[fuse]==1.49.0, azureml-dataset-runtime[fuse]==1.50.0, azureml-dataset-runtime[fuse]==1.51.0, azureml-dataset-runtime[fuse]==1.52.0, azureml-dataset-runtime[fuse]==1.53.0, azureml-dataset-runtime[fuse]==1.54.0, azureml-dataset-runtime[fuse]==1.54.0.post1, azureml-dataset-runtime[fuse]==1.55.0, azureml-dataset-runtime[fuse]==1.56.0, azureml-dataset-runtime[fuse]==1.57.0, azureml-dataset-runtime[fuse]==1.58.0, azureml-dataset-runtime[fuse]==1.59.0, numpy==1.26.4 and pyarrow==18.1.0 because these package versions have conflicting dependencies. Finally, whenever I register another model that is not trained by a Kedro pipeline, but by an Azure ML pipeline, and then follow a similar approach in deploying that, then the deployment and inference works fine. Therefore I think the error has something to do with the training of the model by the Kedro pipeline... Is this familiar to any of you?

Someone will reply to you shortly. In the meantime, this might help:

hi @Quint Van Leeuwen, sorry you're having a bumpy experience. I see a long list with different versions of

azureml-dataset-runtime[fuse]

. do you know who generated those requirements? trying to install

azureml-dataset-runtime[fuse]==1.10.0

and

azureml-dataset-runtime[fuse]==1.11.0

will already be incompatible 🤔

I don't know where those

azureml-dataset-runtime[fuse]

requirements come from. This library is not defined in my conda.yaml file

does it still fail if you try to strip Kedro out of the equation?