A deployment pipeline turns software and model outputs into artifacts that are easier to release. In an ML system, packaging separates training from deployment so training changes have less impact on production delivery. section 10.1.1
Useful pipeline work can include:
- Run unit and integration tests, build artifacts, and add checks such as security testing, performance testing, linting, or deployment to test environments. section 13.3.1
- Package models with configuration files in formats suited to deployment; review validation metrics and error-analysis reports before release. section 10.1.1
- Plan for production failures with monitoring, alerting, incident response guidance, and appropriate operational tools. section 13.3.1
A portable model package
Export a model as ONNX for backend serving or CoreML for iOS, and export its configuration as JSON. Keeping these outputs relatively independent of training helps limit deployment changes when training evolves. section 10.1.1
