Deserialisation of Lifelines models for Azure Managed Online Endpoint

[cjourneay]

I am trying to implement a couple of Lifelines models in an Azure Managed Online Endpoint and I am getting an odd error from the Azure Inference Server module when I try to deserialise the model in the init() call of the scoring file (see here - https://learn.microsoft.com/en-us/azure/machine-learning/concept-endpoints-online?view=azureml-api-2)

On a simple pickle.load() command it raises an exception and throws out the following:

"Can't instantiate abstract class Formula with abstract methods init, differentiate, get_model_matrix, required_variables"

This happens both with a pickle.load() and the joblib equivalent (which is pickle under the hood anyway).

I assume that this is an issue with MS because a) it is their code throwing the error and b) the entire code runs absolutely beautifully on my local machine.

Has anyone successfully deployed Lifelines models to an Azure Managed Online Endpoint? Is there anything that I might be overlooking in the remote environment?

[cjourneay]

For the record I resolved this. Just in case anyone else hits the same deserialisation issues:

1. Check your package versions for consistency across all environments as some packages are incredibly sensitive to even minor version changes - in particular scipy, pandas, and python pickle modules.
2. You may need to dive deep into release notes to understand why - in particular changes around supported versions and deprecations
3. You may need to create custom inference docker images rather than the MS curated versions to overcome these things