Options across the top, criteria down the side, weights agreed before anyone scores anything. The output is not the winning total; it is a visible record of what mattered and how each option was judged, which is what a reviewer can argue with.
The shape
| Criterion | Weight | A: Managed Kafka | B: Self-managed Kafka | C: Cloud pub/sub |
|---|---|---|---|---|
| Ordering guarantees needed | 25% | 5 — per-partition | 5 — per-partition | 2 — no global order |
| Replay / retention (30 days) | 20% | 5 | 5 | 2 — 7 days max |
| Operational burden | 20% | 4 — vendor runs it | 1 — needs 2 FTE | 5 — fully managed |
| Existing team skill | 10% | 3 | 3 | 2 |
| 3-year TCO | 15% | 2 — ₹3.4 cr | 3 — ₹2.6 cr + staff | 4 — ₹1.9 cr |
| Exit / portability | 10% | 4 — open protocol | 5 | 1 — proprietary API |
| Weighted total | 4.05 | 3.65 | 2.75 |
Decision: Option A. Why the number is not the reason: replay for 30 days is a hard requirement from the reconciliation process, and option C fails it outright at 2 — it should have been screened out before scoring rather than allowed to accumulate points elsewhere. Between A and B the totals are close enough to be noise; the deciding factor is that two dedicated operations staff do not exist and are not being funded.
When you produce it
For a contested decision with three or more genuine options and stakeholders who disagree about what matters. Not for a decision that is already obvious — the matrix will be reverse-engineered to justify it and everyone will know.
Who reads it
The review board, who should challenge the weights rather than the scores. Finance and procurement, in a vendor selection. The next architect, who needs to know what was actually being optimised for.
What good looks like
- Weights agreed and published before scoring. This is the entire integrity of the method.
- Hard requirements are screened first as pass or fail, not scored — otherwise an option that cannot meet a mandatory need wins on points elsewhere.
- Each score carries a one-line justification, as above.
- The narrative states the real reason, and says so where it differs from the arithmetic.
- Close totals are treated as ties, not as results.
Common mistakes
- Tuning the weights until the preferred option wins. Visible to everyone, and it destroys trust in every later matrix.
- Scoring mandatory requirements instead of screening on them.
- False precision — a total of 4.05 against 3.65 is not a meaningful margin.
- Presenting the matrix as the decision. It is the working; the judgement is still yours to make and to own.