During our MLOps rotation we usually see some errors coming from LiftWing model services which are difficult to reproduce and follow, due to the absence of the input request.

This task can be connected with this one: https://phabricator.wikimedia.org/T416384 .

One the one hand, we need to avoid overloading logstash with extra noise by logging everything, but on the other hand we need to find a way to debug and reproduce the errors based on the incoming request which caused the error.

This task is pretty generic for now and it will be enhanced with much more details for each model.