MLOps operationalizes machine-learning systems through tools and practices for building and maintaining pipelines. It complements ML system design, which connects technical work with software engineering, project management, product management, and leadership. section 1.4 section 10.2
A practical MLOps view follows connected operational areas:
- Training pipelines coordinate data and model-building steps. section 10.2
- Inference and deployment pipelines move trained systems toward use. section 10.2
- Monitoring services support the operating ML system. section 10.2
- Pipeline platforms should manage dependencies, reproduce results from parameters and versions, connect to compute resources, and store artifacts. section 10.2
Simple pipeline implementation
A complete training pipeline can begin as Python scripts connected by shell scripts. Specialized ML tools add abstractions and impose coding conventions that can improve consistency and maintainability, but they are not always necessary. section 10.2
Figure 10.3 Using proper frameworks for a training pipeline is usually a good practice, although sometimes it is enough to keep things simple.
Compare the value of adopting a dedicated training-pipeline framework with keeping an implementation simple. section 10.2
