This SLO and error budget worksheet gives service owners and SRE teams a repeatable way to choose reliability targets, calculate allowed failure, define alert thresholds, and agree on what happens when reliability work consumes the budget.
Overview
A service-level objective (SLO) is a reliability target measured over a defined time window. An SLO might state that 99.9% of valid requests should succeed during a rolling 30-day period, or that 99% of eligible requests should complete within a specified latency threshold. The measurement used to evaluate that target is the service-level indicator (SLI).
The error budget is the amount of unreliability permitted by the SLO. It converts a target into an operating constraint that product, engineering, and operations teams can use when making trade-offs. If a team has used most of its budget, it may choose to reduce risky releases, prioritize remediation, or investigate a recurring failure mode. If substantial budget remains, the team has more room to deliver planned changes, provided other risks are understood.
A useful SLO is specific enough to measure and important enough to influence decisions. Avoid choosing targets only because a rounded number sounds impressive. A target that no team monitors, owns, or acts on is documentation rather than an operational control.
Worksheet summary
| Field | Decision to record |
|---|---|
| Service and user journey | What service or customer-critical action is covered? |
| SLI | Which measurable events represent success and failure? |
| SLO target | What percentage of valid events must succeed? |
| Window | Which calendar or rolling period is evaluated? |
| Exclusions | Which events are outside the service team's control, and why? |
| Error budget policy | What actions follow when the budget is low or exhausted? |
| Owner | Who maintains the query, dashboard, alert, and policy? |
How to estimate
Start with the user-visible outcome rather than an infrastructure metric. CPU utilization, pod restarts, and database connections can be useful diagnostic signals, but they are not automatically good SLIs. Ask what a user considers successful: completing a checkout, receiving an API response, loading a page, or processing a message within an acceptable time.
1. Define the eligible event
Write the event in a way that a query can identify consistently. For an API, the numerator might be requests that return an acceptable response within the latency threshold, while the denominator is all valid requests. For a batch service, the denominator could be scheduled jobs and the numerator jobs completed correctly within the required time.
Document exclusions explicitly. Examples might include deliberately rejected requests that are invalid according to the API contract, test traffic, or an upstream dependency outside the agreed service boundary. Exclusions should not be used to hide failures that real users experience. If an exclusion is difficult to identify reliably, treat that as a measurement problem rather than silently removing the event.
2. Select the target and calculate the budget
The basic formulas are:
- Allowed failure percentage = 100% − SLO target.
- Error budget in events = eligible events × allowed failure percentage.
- Error budget in time = measurement-window duration × allowed failure percentage.
Choose the target by considering user expectations, business impact, dependency limits, historical performance, and the team's ability to detect and correct failures. A higher target is not automatically better: it permits less failure and can make normal changes harder to operate unless the system and process can support it.
3. Define burn rate and alerts
Burn rate describes how quickly the service is consuming its budget compared with the rate allowed by the SLO. A simple form is:
Burn rate = observed error rate ÷ allowed error rate.
For example, with a 99.9% SLO, the allowed error rate is 0.1%. If the observed error rate is 2%, the burn rate is 20 times the permitted rate. The duration of the observation window matters: a short, severe failure may require immediate response, while a slower trend may be better suited to a ticket or planning discussion.
Use alert windows that match the response you expect. A page should generally indicate a credible need for prompt human action; a lower-severity notification can track gradual budget consumption. Pair a short window with a longer window when possible so that a brief spike does not create unnecessary escalation while a sustained problem is not hidden by a long average. Treat these as starting points and validate them against incident history and alert fatigue.
Inputs and assumptions
Complete the following worksheet before publishing an SLO:
- Service: Record the service name, environment, and user journey covered.
- Population: Define the traffic, tenants, regions, or job types included.
- SLI query: Identify the source metric, numerator, denominator, aggregation, and treatment of missing data.
- Target: Record the success percentage and whether it applies to availability, latency, correctness, freshness, or another dimension.
- Window: Choose a calendar or rolling period and record the timezone or evaluation convention where relevant.
- Ownership: Name the service owner and the team responsible for maintaining dashboards and alert rules.
- Policy: Define actions for normal operation, low remaining budget, exhausted budget, and disputed measurement.
For latency SLOs, define the threshold and population precisely. “Fast” is not queryable; “eligible requests completed within the agreed threshold” is. Decide whether the objective applies to all requests or a defined percentile-style success ratio, and ensure the telemetry can distinguish completed requests from abandoned or missing observations.
For availability, confirm that the monitoring path represents the user's experience. A healthy process metric does not prove that requests are succeeding at the edge, through authentication, or across dependent systems. When a service depends on another team, document whether the dependency failure is included in the service's SLO, tracked separately, or represented through a composite user journey.
Worked examples
Availability over a 30-day window
Assume a service has a 99.9% availability SLO over a 30-day window. Using a 30-day approximation of 43,200 minutes:
- Allowed failure percentage: 100% − 99.9% = 0.1%.
- Error budget: 43,200 × 0.001 = 43.2 minutes.
The team therefore has approximately 43.2 minutes of measured unavailability in the window before the budget is exhausted. If 18 minutes have been consumed, the remaining budget is approximately 25.2 minutes. The worksheet should also record how partial minutes, maintenance, excluded traffic, and measurement gaps are handled; the arithmetic alone does not settle those policy questions.
Request success by event count
Assume the same 99.9% target applies to 100,000 eligible requests. The allowed failure percentage is 0.1%, so the event budget is 100,000 × 0.001 = 100 failed requests. If 35 requests have failed, 65 failures remain within the measured window. A dashboard can show both the remaining count and the percentage consumed so that the result is understandable during planning or an incident.
Latency as a success ratio
Suppose a team defines a latency SLO requiring eligible requests to complete within a stated threshold. With 50,000 eligible requests and a 99% target, up to 500 requests may exceed the threshold before the budget is exhausted. This is different from saying that the average response time must remain below the threshold: a success-ratio SLO counts each eligible event according to whether it met the defined condition.
When a budget is exhausted, the policy should be specific. Possible actions include pausing non-essential production changes, assigning an owner to a corrective work item, reviewing recent deployments, or requiring a reliability review before a higher-risk launch. Link the policy to your incident process, such as the SRE incident response runbook, rather than creating a separate response path that nobody checks.
When to recalculate
Recalculate the worksheet during quarterly planning, launch readiness reviews, reliability incidents, and error-budget policy checks. Revisit it sooner when the service's traffic pattern, architecture, dependency boundary, user journey, or monitoring implementation changes.
Review the SLI query whenever telemetry fields or dashboard logic change. A deployment that changes request routing, retries, sampling, status codes, queue semantics, or regional traffic can alter the meaning of the measurement without changing the SLO text. Validate the query against known successful and failed scenarios before relying on the result.
After an incident, compare the measured budget consumption with the observed customer impact. If the budget was nearly untouched despite a serious user-facing problem, the SLI may be incomplete. If routine, low-impact events consume the entire budget, the event definition, target, or service boundary may need review. Do not simply raise the target or add exclusions to make the chart look healthier.
Keep the worksheet alongside the service's dashboard and ownership documentation. For teams operating Kubernetes or other shared platforms, connect the SLO to the deployment strategy and rollback conditions described in the Kubernetes deployment strategy guide. For services onboarding to a platform, include the SLI query, alert ownership, and error-budget policy in the platform onboarding checklist.
Action checklist: choose one user-critical journey, write its numerator and denominator, select a defensible target, calculate the time or event budget, define alert windows, assign an owner, and record the response when the budget is low or exhausted. Re-run those calculations whenever the underlying traffic, architecture, or measurement assumptions change.