Bento

A Bento is a deployment-ready archive for an ML application. It packages a trained model with its serving code, dependencies, and configuration so the application can be distributed and deployed as one unit. It replaces the ad hoc task of assembling those pieces separately for each deployment; it is not itself a running container or service. section 6.1 section 10.5

To package an income-classifier API:

• Import the trained model from MLflow into the local BentoML model store. Check its name and tag with `bentoml models list`. section 6.1.1
• Define the BentoML service and its runner in `service.py`. The `/predict` endpoint retrieves online features from Feast, prepares the model input, runs inference, and returns a prediction. section 6.1.1
• Define `bentofile.yaml` with the service entry point, files to include, Python requirements, and Docker environment settings. The build configuration determines what accompanies the service. section 6.1.1
• Run `bentoml build -f bento/bentofile.yaml`. Inspect the resulting Bento with `bentoml list`, then test it locally before pushing it to a remote Bento store. section 6.1.2

service: "service:svc"
labels:
  owner: ml-engineering-team
include:
  - "service.py"
  - "production.env"
  - "src/*.py"
  - "requirements.txt"
  - "*.env"
python:
  requirements_txt: requirements.txt
docker:
  env:
    - ENV_NAME=local
    - AWS_ACCESS_KEY_ID=minio
    - AWS_SECRET_ACCESS_KEY=minio123

The income-classifier `bentofile.yaml` shows the service entry point alongside included source files, a requirements file, and environment settings. Check this manifest when a build lacks code or configuration the service needs. section 6.1.1

A successful build receives a tag such as `income_classifier_service:2umhoidopolz3ktg`, identifying that version of the service. `bentoml list` reports application size separately from model size. Containerization is a subsequent step: `bentoml containerize` can create a Docker image, while Yatai can build an image after a Bento is pushed for deployment. section 6.1.2 section 6.1.3

For a Kubernetes deployment, push the tested Bento to Yatai and verify both the Bento and model in their remote stores. Select the Bento and tag when creating a deployment, then configure replicas and CPU and memory resources. Yatai launches an API Server to receive requests and Runner pods to perform inference; these can be scaled independently. Verify the pods and exercise the deployed `/predict` endpoint. section 6.1.2 section 6.1.3