Before choosing a model or tool, state the problem the system must solve. Ask what business goal it serves, what it may cost, and whether its likely output justifies that cost. A sophisticated implementation cannot rescue a poorly defined or unsolvable problem. section 2.0
Machine learning system design is the architectural plan that connects an ML system’s building blocks. MLOps supplies tools for building and maintaining the system; design provides a blueprint for how the pieces fit together, where bottlenecks may arise, and how the system can remain flexible as it grows. It draws on software engineering as well as ML, product management, project management, and leadership. section 1.1.1 section 1.4
Treat design as a set of decisions to revisit, not a demand to implement every possible component:
- Articulate the problem before entering the solution space; consider whether it is solvable with available technology. section 2.0
- Clarify purpose, budget, deadlines, expected value, risks, and limitations before scoping the project. section 1.4 section 2.0
- Use a roadmap and checklists to consider each part of the system, including why and when a technique or tool is useful. Not every checklist item needs to be implemented. section 1.2 section 1.4
When a blueprint matters
A shed might be built without a detailed plan; a house or skyscraper cannot sensibly be approached the same way. Likewise, a simple ML project may need little formal design, while a larger system benefits from an explicit plan for its components and their connections. section 1.1.1
Do not assume every proposed problem deserves an ML build. Some must wait for better technology; others are ill-posed, such as trying to devise an algorithm that reliably beats casino roulette. Establish that distinction before committing to a solution. section 2.0
