- When you want to quickly release the predictor in the simplest architecture.
The web single pattern is an architecture that packs all the artifacts and code for the prediction model in a web server. Since the single server REST (or GRPC) interface, preprocess, and trained model are in one place, you can create and deploy them as a simple predictor.
If you want to deploy multiple replicas, you need to deploy them behind a load balancer or proxy. In case you are using GRPC for the interface, you need to consider client side load balancing or layer-7 load balancer.
To build your model into a web server, you can adopt either model-in-image pattern or model-load pattern.
- Able to use one programming language, such as Python, for the web server, preprocess and inference.
- Easy to manage because of it's simplicity.
- Easier troubleshooting.
- Minimal time-to-productionize post model training.
- It is usually suggested to start your production architecture with deploying your models in the web single pattern in a synchronous fashion.
- Since all components are packed in a server or a docker image, applying a small patch will require updating the whole image
- Updates will also require a service deployment, which in bigger organizations requires one to follow a hefty SDLC
- Update and maintenance procedure for each component.
- Scale change management of the web server.