In one line

Features say what a system does; quality attributes say how well it does it. Architecture is mostly about choosing which qualities matter most and paying for them.

The idea

Two systems can have identical features and completely different architectures, because they need different qualities. The five that matter most for the roadmap:

AttributeQuestion it answersTypical measure
ScalabilityCan it handle more load by adding resources?Throughput (images/hour), latency at peak
ReliabilityDoes it keep working correctly when things fail?Availability %, error rate, successful pipeline runs
MaintainabilityCan people understand and change it safely?Lead time for changes, change failure rate, onboarding time
SecurityIs it protected against misuse and data leaks?Findings, least-privilege coverage, secrets in Key Vault
CostIs it affordable to run and grow?Cost per 1,000 images, GPU-hours per model

Others worth recognising: performance (latency), availability, observability, testability, portability, compliance.

How it works

Make them measurable

“It should be fast” is not a requirement. Write a quality attribute scenario:

When 200 images per minute arrive (stimulus) during a warehouse shift peak (environment), the ingestion stage (artefact) validates and stores them (response) with p95 under 5 seconds and zero lost files (measure).

That sentence can be tested, monitored and argued about. See 01 Requirements for the full format.

They pull against each other

flowchart LR
  SC[Scalability] <-->|more parts to manage| M[Maintainability]
  R[Reliability] <-->|redundancy costs money| C[Cost]
  SE[Security] <-->|more checks, more friction| P[Performance]
  SC <-->|more instances, more spend| C

You can’t maximise all of them. A good architect names the top two or three for a system and accepts weaker results elsewhere. 09 Trade-off Analysis shows how to compare options against them.

Microsoft’s version: Well-Architected pillars

The Azure Well-Architected Framework groups them into five pillars: Reliability, Security, Cost Optimization, Operational Excellence, Performance Efficiency. It’s worth learning because onshore architects and AZ-305 both use this language.

On the platform

For a machine-vision platform, a sensible ranking might be:

  1. Reliability of data: no image or label silently lost or corrupted. A bad dataset poisons every model trained on it.
  2. Maintainability: the pipeline changes constantly as models and annotation needs evolve.
  3. Cost: GPU training and annotation are the big bills; storage tiers matter at 3D data volumes.
  4. Scalability: usually bursty (a new site sends a big batch), not constant.
  5. Security: customer warehouse images may be sensitive; access control on Blob containers and secrets matter.

These map directly onto the platform KPIs in the roadmap: turnaround time (performance/scalability), annotation cost per batch (cost), evaluation pass rates (reliability of models), deployment frequency (maintainability).

Trade-offs

  • Strong consistency and reliability → more coordination → lower throughput (see 08 CAP Theorem and Consistency).
  • Hot storage tiers → fast reads, higher cost; archive tiers → cheap, slow to retrieve.
  • More security checks in CI → safer, slower pipelines.

Common mistakes

  • Treating qualities as an afterthought (“we’ll make it scale later”).
  • Unmeasurable goals: “highly available”, “secure”, “fast”.
  • Optimising a quality the business doesn’t care about (five-nines availability for an internal batch job).

Check yourself

Apply it

  • Write three quality attribute scenarios for the real platform (one each for reliability, cost and turnaround time).
  • Rank the top three qualities for the platform and check the ranking with a tech lead. Disagreement here is valuable.

Go deeper

Related: 01 Requirements · 09 Trade-off Analysis · 08 Architecture Decision Records