In one line
Functional requirements describe behaviour (“import labels from V7”). Non-functional requirements (NFRs) describe qualities and constraints (“import 10,000 labels in under 2 minutes, with no data loss”). Architecture is driven mostly by the second kind.
The idea
| Functional | Non-functional | |
|---|---|---|
| Answers | What does it do? | How well, under what limits? |
| Example | Users can search images by site and date | Search returns in < 500 ms p95 for 50M images |
| Verified by | Feature tests, acceptance criteria | Load tests, chaos tests, monitoring, audits |
| Changes architecture? | Sometimes | Almost always |
NFR categories map to 04 Quality Attributes: performance, scalability, availability, reliability/durability, security, maintainability, cost, compliance, observability.
Constraints are a third kind worth listing separately: things you can’t choose (“must run on Azure”, “data stays in the EU”, “team of six”, “go-live by March”).
How it works
Start every design by clarifying both
flowchart TD Q[Problem statement] --> F[Functional:<br/>who does what?] Q --> N[Non-functional:<br/>how much, how fast,<br/>how reliable?] Q --> C[Constraints:<br/>tech, budget, time,<br/>compliance] F & N & C --> S[Scope for v1<br/>+ explicit out-of-scope]
Make NFRs measurable
Weak → strong:
| Weak | Strong |
|---|---|
| ”Fast" | "Upload accepted in < 2 s p95 for files up to 20 MB" |
| "Highly available" | "Ingestion API 99.9% monthly availability (≈ 43 min downtime)" |
| "Scalable" | "Handles 5× current daily volume without redesign" |
| "Secure" | "Only managed identities access Blob; no secrets in code; images encrypted at rest" |
| "Cheap" | "Storage cost < €X per TB per month after 90 days” |
Useful words: p50/p95/p99 latency, throughput, RPO (how much data can you afford to lose) and RTO (how long can you be down), SLO (internal target) vs SLA (contractual promise).
Availability cheat sheet
| Availability | Downtime per month |
|---|---|
| 99% | ~7.3 hours |
| 99.9% | ~43.8 minutes |
| 99.99% | ~4.4 minutes |
Every extra nine roughly multiplies cost and complexity. Ask: what does an hour of downtime actually cost?
On the platform
Example requirement set for an image ingestion capability:
Functional
- Robots/cameras upload images with site, robot ID and timestamp metadata.
- Each image is validated (format, resolution, not corrupt, not duplicate).
- Valid images are searchable by site, date and robot.
- Invalid images are quarantined with a reason.
Non-functional
- Peak 300 images/min per site during commissioning; 20 sites.
- Zero silent data loss (durability over latency).
- Validation result available within 60 s of upload (p95).
- Raw images retained 2 years; moved to cool tier after 30 days.
- Ingestion availability 99.5% (uploads can retry on the robot side).
Constraints: Azure, Python, existing MongoDB, data residency in the EU.
These numbers are placeholders
The exercise is to replace them with real figures from the team. Getting real numbers is often the most useful outcome of the exercise.
Common mistakes
- Jumping to a solution (“we need Kafka”) before agreeing on volumes and latency.
- NFRs that nobody measures after go-live.
- Gold-plating: designing for 99.99% when 99.5% is fine.
- Forgetting operational requirements: backups, monitoring, on-call, cost reporting.
Check yourself
Give one functional and one non-functional requirement for model deployment.
Functional: “A registered model can be promoted to production with one approval.” Non-functional: “Promotion completes within 10 minutes and can be rolled back within 5.”
What are RPO and RTO?
RPO (Recovery Point Objective): the maximum acceptable data loss, measured in time. RTO (Recovery Time Objective): the maximum acceptable time to restore service.
What’s the difference between an SLO and an SLA?
An SLO is an internal reliability target. An SLA is a promise to a customer, usually with penalties. SLOs should be stricter than SLAs.
Apply it
- Write functional + non-functional requirements for one platform capability (ingestion is the easiest). Get real numbers from the team.
- Turn three of the NFRs into draft platform KPIs (month-1 deliverable).
Go deeper
Related: 02 Back-of-Envelope Estimation · 04 Quality Attributes · 09 Trade-off Analysis