I recently went through the workflow again and thought it might be better to implement our APIs in stateless manner, i.e., instead of tracking a current_model at the backend, we can instead pass in a model_id parameters in all model APIs - specifying which model to perform this operation on.
Pros:
Managing current_model state is tricky and error prone. Difficult to keep frontend in sync with the backend
current_model design has poor extensibiliy, and tightly couples with our current architecture.
We can only track one model at a time
Cons:
Given our current architecture which assumes user deploy the app in their local machine, it's difficult to run these APIs with different model (e.g., what if the user is making interleaving and frequent calls model-1 and model-2, does the back end pre-load them, or load just in time on demand?)
I recently went through the workflow again and thought it might be better to implement our APIs in stateless manner, i.e., instead of tracking a
current_modelat the backend, we can instead pass in amodel_idparameters in all model APIs - specifying which model to perform this operation on.Pros:
current_modelstate is tricky and error prone. Difficult to keep frontend in sync with the backendcurrent_modeldesign has poor extensibiliy, and tightly couples with our current architecture.Cons: