Define a request-based SLI and SLO, calculate an error budget and write an incident-review rule for a synthetic service.
A fictional library API serves 100,000 eligible requests in a 30-day window. Its objective is 99.9% successful requests. The sample contains 80 failed requests. Success is defined by a documented response rule; excluded health checks are not in the denominator.
| Measure | Value |
|---|---|
| Eligible requests | 100000 |
| SLO percent | 99.9 |
| Failed requests | 80 |
| Window days | 30 |
An SLI needs a population, event and measurement rule. Choose something related to the user experience rather than a convenient internal count. Define eligible requests and success consistently. The case uses a request-based ratio, so translating it directly into minutes of downtime would change the measurement model.
Write the SLI definition and exclusions.
SLI specification
Record your result and evidence here or in a separate file.
The denominator is eligible requests, not every log entry.
Explain every exclusion and keep it stable across the window.
The allowed bad-event fraction is one minus the objective. Multiply it by eligible requests to obtain the budget for this sample window. The actual success rate and budget consumption answer related but different questions. Keep the units visible and avoid rounding before the final calculation.
Calculate actual success, allowed failures and remaining budget.
Budget sheet
Record your result and evidence here or in a separate file.
Success is 99.92%; budget is 100; 80 are consumed and 20 remain.
Use 0.001, not 0.1, as the allowed failure fraction.
An error-budget policy connects reliability evidence with a decision process. It should describe who reviews risk, what work may be prioritized and how exceptions are recorded. A policy is not an automatic ban on every change. This fictional exercise does not authorize operations on any real service.
Draft a review rule for budget exhaustion and a documented exception route.
Budget policy
Record your result and evidence here or in a separate file.
The policy identifies an owner and evidence, not just a red status.
Replace vague be more careful instructions with a specific review action.
An alert needs enough traffic and a useful time window. A single failure in a tiny sample can create a dramatic rate without justifying the same response as a sustained incident. This path does not derive a full burn-rate alert system; it asks for a proportionate evaluation plan and an explicit data-sufficiency rule.
Compare 1 failure in 10 requests with 1,000 in 10,000 and describe the context needed.
Alert design note
Record your result and evidence here or in a separate file.
The note distinguishes equal 10% rates from very different event volumes.
Show counts beside rates and state the response context.
A reliability review combines the measured objective, budget state, incidents and improvement actions. Avoid claiming the service is reliable in every dimension because one ratio passes. Latency, data correctness and availability may need separate measures. The supplied case contains no live monitoring evidence.
Write a one-page monthly review and two next measurement questions.
Reliability report
Record your result and evidence here or in a separate file.
The report limits its conclusion to the supplied request-success measure.
Name at least one user-experience dimension not measured.
Deliver the SLI contract, budget calculation, policy, alert note and review. Recalculate with 120 failures and explain the policy response without inventing a live incident.
Request ratios are not silently converted to downtime, and the denominator remains explicit.
| Criterion | Points | Evidence |
|---|---|---|
| Evidence and correctness | 30 | Calculations, source use and results match the supplied case; limitations are honest. |
| Required artifacts | 30 | All five stage outputs and the capstone handoff are present and inspectable. |
| Reasoning and decisions | 20 | Choices, assumptions, alternatives and the case-specific critical condition are explained. |
| Reproducibility | 10 | A reviewer can follow the stated steps and distinguish expected from observed results. |
| Clear communication | 10 | Labels, units, structure and evidence references make the work easy to review. |
Score at least 80/100, supply every required artifact, and satisfy the path-specific critical condition. Rework missing evidence before marking complete.
Allowed failures are 100,000 × (1 − 0.999) = 100. Actual success is 99,920 / 100,000 = 99.92%. The case consumes 80% of its error budget and leaves 20 failed requests of allowance for the same fixed window and population.
This is an SLO design and calculation exercise, not a monitored production service or contractual availability guarantee.