How far does model size and lag impact distributed inference ?

[anvil]

Hello !

I was wondering how impactful a model's size of inference lag was in a distributed manner.

With tools like Pandas Iterator UDFs or mlflow.pyfunc.spark_udf() we can make it so models are loaded only once per worker, so I would tend to say that minimizing inference lag is more important than minimizing size, since size will impact us once per model whereas lag will impact us once per observation.

I would also say that the impact is even greater with ensemble models where several models - with their own lag - each need to infer once per observation.

Is this assumption correct ?

Thank you !

[youssefmrini]

Your assumption that minimizing inference lag is more important than minimizing the size of the model in a distributed setting is generally correct.

In a distributed environment, models are typically loaded once per worker, as you mentioned, which means that the impact of model size is limited to the initial loading of the model. However, inference lag occurs every time an observation is processed, which can have a significant impact on the overall performance of the system.