Cost Report

Every design decision has a price tag. A cache that saves 90% of database reads, a CDN that absorbs the on-sale, three availability zones instead of one, an agent that calls the LLM five times per question — each one shows up on the monthly bill. In the real world, "it scales" is only half an answer; the other half is "and it costs US$ X per month, or US$ Y per ticket sold". Milestone 8 asks your team to price its design honestly, connect the cost to the measurements of the load-test campaign, and say where the money would be better spent.

This milestone builds on the FinOps part of Class 11 — Cloud-native, on your M1 estimates and on the load-test campaign. Due in week 18, weight 10% of the project.

What to deliver

• A cost report page in the docs, following the template.
• The calculator export (shareable link, PDF or CSV) with the date and the region used — prices change, so the export is the evidence.
• A small script or spreadsheet with the formulas (committed), so that the numbers can be recomputed.
• The final update of the design doc (with a changelog of what changed since M1).
• The tag m8.

Acceptance criteria

• (mandatory) A bill of materials covering every component of your design doc (compute, databases, cache, broker, load balancer, NAT, egress/CDN, storage, observability, and LLM tokens for the AI option), each with quantity, size, hours and the price used.
• (mandatory) Every price comes from an official pricing calculator or pricing page, with date and region; no price is invented. Assumptions (hours per month, GB transferred, utilization, retention) are listed.
• Unit economics: cost per 1,000 requests and cost per business unit of your option.
• Cost of the environment you actually used (laptop, CI minutes, any cloud account — with the bill if a cloud was used).
• Sensitivity: cost at ×10 traffic, provisioned-for-peak vs autoscaled, and with commitments/spot (reserved instances, savings plans, committed-use discounts).
• A service-model comparison for at least one component (e.g. Kafka self-managed on Kubernetes vs a managed Kafka service vs a serverless queue): cost, operational effort, limits.
• ≥ 3 saving levers, each tied to a measurement from M6 (e.g. "hit ratio 95% → database can be one size smaller") with an estimated saving.
• AI usage section (see the policy).

Step by step

1. Pick one cloud and its calculator

Use one provider for the whole estimate (you may compare a single component across providers as an extra). Map each palette component of your design to a concrete service:

2. Size each component from your numbers

Start from the M1 estimates and the M6 measurements, not from guesses: the capacity per replica you measured, the peak of the design target, a target utilization (≈ 0.6–0.7), the hit ratio of the cache, the GB that leave the cloud (egress is often the surprise), retention of logs and metrics, and — for the AI option — tokens per request × requests.

3. Visualize where the money goes

A stacked chart of the bill by category makes the dominant costs obvious. The chart below uses illustrative placeholder prices — not real quotes — only to show the shape of the analysis; your report must use the prices of your calculator export.

4. Unit economics

Divide by the unit your business cares about. For BoxOffice, cost per ticket sold is the number a product manager understands; for chat, cost per 1,000 messages; for video, cost per hour watched; for the assistant, cost per conversation (with tokens as the main driver). Compare it with what the unit is worth (a ticketing fee, an ad impression) — an architecture that costs more per ticket than the fee is not a business.

5. Service models

Pick at least one component and compare how you would run it:

Put the effort in hours per month next to the bill: an engineer's time is the largest hidden cost of self-managed components.

6. Saving levers backed by measurements

Every lever must point to a measurement: "M6 shows the catalog cache at 95% hit ratio, so the database can drop one instance size (−US$ …)", "CPU of notification averages 8% — move it to serverless or a smaller node pool", "the semantic cache answered 30% of questions — LLM tokens −30%", "logs are 60% of observability cost — sample debug logs and shorten retention". Also consider commitments (1-year reserved capacity or savings plans), spot/preemptible nodes for stateless workers, and the egress saved by the CDN.

Templates

# Cost report — <team / system>

## 1. Scope and assumptions
Cloud and region: ... · Calculator export: <link/file>, date ...
Design target: ... req/s peak, ... average; 730 h/month; egress ... GB/month; retention ...

## 2. Bill of materials (design target)
| Component (palette id) | Service | Size | Qty | Hours/month | Unit price (source) | US$/month |
|-|-|-|-|-|-|-|
| app (catalog) | ... | 2 vCPU / 4 GB | ... | 730 | ... | |
| sql | ... | | | | | |
| **Total** | | | | | | |

## 3. Unit economics
| Unit | Monthly volume | Cost per unit |
|-|-|-|
| 1,000 requests | | |
| ticket / message / ride / view / conversation | | |

## 4. Environment actually used
Laptop / CI minutes / cloud account (bill attached): ...

## 5. Sensitivity
| Scenario | US$/month | Notes |
|-|-|-|
| Baseline (design target) | | |
| ×10 traffic | | which items grow linearly, stepwise, stay fixed |
| Provisioned for peak vs autoscaled | | |
| With commitments / spot | | |

## 6. Service-model comparison: <component>
| Option | US$/month | Effort (h/month) | Limits | Choice |
|-|-|-|-|-|

## 7. Saving levers
| Lever | Evidence (M6 measurement) | Estimated saving | Risk |
|-|-|-|-|

## 8. AI usage

# Relatório de custos — <time / sistema>

## 1. Escopo e premissas
Nuvem e região: ... · Exportação da calculadora: <link/arquivo>, data ...
Meta de design: ... req/s no pico, ... na média; 730 h/mês; tráfego de saída ... GB/mês; retenção ...

## 2. Lista de materiais (meta de design)
| Componente (id da paleta) | Serviço | Tamanho | Qtd | Horas/mês | Preço unitário (fonte) | US$/mês |
|-|-|-|-|-|-|-|
| app (catalog) | ... | 2 vCPU / 4 GB | ... | 730 | ... | |
| sql | ... | | | | | |
| **Total** | | | | | | |

## 3. Economia unitária
| Unidade | Volume mensal | Custo por unidade |
|-|-|-|
| 1.000 requisições | | |
| ingresso / mensagem / corrida / visualização / conversa | | |

## 4. Ambiente realmente usado
Notebook / minutos de CI / conta na nuvem (fatura anexada): ...

## 5. Sensibilidade
| Cenário | US$/mês | Observações |
|-|-|-|
| Base (meta de design) | | |
| ×10 de tráfego | | o que cresce linearmente, em degraus, fica fixo |
| Provisionado para o pico vs autoscaling | | |
| Com compromissos / spot | | |

## 6. Comparação de modelos de serviço: <componente>
| Opção | US$/mês | Esforço (h/mês) | Limites | Escolha |
|-|-|-|-|-|

## 7. Alavancas de economia
| Alavanca | Evidência (medição do M6) | Economia estimada | Risco |
|-|-|-|-|

## 8. Uso de IA

Rubric

Levels as defined in How the exercises are graded.

FAQ