MLOps Foundations and Solution Scoping
Domain 1 tests how well you frame an ML initiative before tooling: lifecycle stages, MLOps maturity, stakeholder alignment, and feasibility. The exam expects you to connect a business outcome to a measurable ML objective and to an operating constraint, not just name a platform.
Lifecycle and maturity
- MLOps maturity commonly moves from manual, notebook-driven level 0 to automated training pipelines at level 1 and full CI/CD/CT at level 2.
- A solution scope must name the user, decision, label source, latency/throughput target, data refresh cadence, and acceptance metric.
- Build-vs-buy depends on differentiators, compliance, team skills, and total cost; managed services are valid when the model is not the product.
Scoping rules
Translate the business KPI to a model metric with an explicit threshold. For defect detection, scrap rate is the business KPI; recall at a fixed precision is the model metric; edge latency under 50 ms is the constraint. Confirm label availability and base rate before promising accuracy.
Worked example
A factory wants visual defect detection. Business goal: cut scrap 15%. ML goal: recall ≥0.95 at precision ≥0.90 on a held-out line. Constraint: 30 FPS on Intel CPUs at the edge. Reason: low recall misses defects and raises scrap; low precision causes false rejects and rework. Scope a pilot on one line, one product family, 8 weeks, with OpenVINO inference and a human review queue. Success requires both quality and p95 latency; otherwise the pilot fails.
Traps
- Starting with a toolkit instead of a decision and metric.
- Scoping without checking label quality, base rate, or feedback latency.
- Treating a proof of concept as production-ready.
- Ignoring inference hardware and operator constraints until deployment.