Use this SLO and error budget review template to turn reliability data into consistent release decisions. It explains how to choose useful service-level indicators, calculate an error budget, interpret burn rates, distinguish customer impact from technical symptoms, and document whether to release, pause, or invest in recovery work.
Overview
A service-level objective (SLO) defines the reliability target for a service over a stated time window. An SLO normally measures a user-visible outcome such as successful requests, acceptable latency, or completed jobs. The error budget is the amount of unreliability the objective permits during that window.
For an availability SLO, the basic calculation is:
Error budget percentage = 100% − SLO target
If a service has a 99.9% availability target for a 30-day review period, its budget is 0.1% of eligible events or time, depending on how the SLI is defined. That example is illustrative: teams should use the target, event model, and review window that match their service contract and customer experience.
The budget is not a reward for creating outages, nor is it a reason to release without judgment. It is a decision signal. When the budget is healthy, the team may have room for planned change. When the budget is being consumed quickly, the team should reduce avoidable risk, prioritize recovery, or pause changes that could worsen customer impact.
Review the SLO alongside deployment history, incident records, support reports, and important customer journeys. A dashboard can show a technical symptom while missing a degraded workflow. Conversely, a noisy metric can suggest failure when customers are not materially affected.
How to estimate release room and burn rate
Begin each review with four values: the SLO target, the review-window length, the amount of budget already consumed, and the remaining time in the window. For event-based SLIs, calculate the ratio of bad events to total eligible events. For time-based availability, calculate unavailable time against the total eligible time. Keep the denominator visible; a percentage without its eligible population is difficult to interpret.
A simple remaining-budget calculation is:
Remaining budget = total budget − consumed budget
Burn rate compares the observed rate of budget consumption with the rate that would consume the entire budget exactly by the end of the SLO window:
Burn rate = observed budget consumption rate ÷ planned budget consumption rate
A burn rate of 1 means the current rate would use the complete budget by the end of the window. A rate above 1 means the team is consuming budget faster than planned. A rate below 1 indicates slower consumption, but it does not prove that a release is safe.
Use more than one time window. A short window can detect a severe incident quickly, while a longer window shows whether the service is drifting over time. The exact thresholds should reflect your service, traffic volume, alert tolerance, and response capacity.
| Observed condition | Suggested review interpretation | Typical action |
|---|---|---|
| Burn rate near or below 1 | Consumption is broadly aligned with the plan | Review the change risk and proceed only with normal safeguards |
| Moderately above 1 | The budget is being consumed faster than intended | Increase scrutiny, reduce change scope, and assign corrective work |
| High burn rate on a short window | Recent customer impact may be significant | Investigate immediately, stabilize the service, and consider a temporary release pause |
| Budget exhausted | The SLO target has been missed for the review period | Follow the error-budget policy and prioritize reliability recovery |
Do not treat these rows as universal thresholds. A low-volume service may need event counts and absolute impact alongside percentages. A high-volume service may need separate views by region, endpoint, tenant, or customer journey.
Inputs and assumptions
Complete this worksheet before deciding whether to release:
- Service: Name the user-facing service and owning team.
- Review window: State the start and end dates, such as a rolling period or planning cycle.
- SLI: Define exactly what is measured, including eligible events and exclusions.
- SLO target: Record the target and whether it covers availability, latency, freshness, correctness, or completion.
- Budget: Convert the target into an event, time, or latency allowance that the team can inspect.
- Consumption: Record total eligible events, bad events, and the resulting budget usage.
- Burn rate: Calculate it for at least one short and one longer window.
- Customer impact: Identify affected journeys, regions, tenants, or priority functions.
- Change risk: Record deployment size, rollback readiness, dependency changes, and traffic exposure.
- Mitigations: List feature flags, staged rollout steps, capacity changes, or support communication.
Define exclusions carefully. Planned maintenance, synthetic checks, internal traffic, and dependency failures may be treated differently, but exclusions can hide real customer impact if they are broad or difficult to audit. Every exclusion should have an owner and a reason.
Latency SLOs require particular care. An average can conceal slow requests affecting a smaller but important group of users. Consider a percentile or another distribution-aware measure, then verify that the query uses the correct route, status, region, and time range. For asynchronous systems, measure completion or freshness rather than applying a request-availability model that does not represent the user experience.
For platform teams, connect the review to deployment and infrastructure records. A Kubernetes rollout, resource-limit change, provider upgrade, or infrastructure-as-code modification should be traceable to the period in which the SLI changed. The supporting runbooks for Kubernetes resource requests and limits and Terraform provider version constraints can help teams connect operational changes with reliability review evidence.
Worked examples
Example 1: Availability and a planned release
Assume an illustrative service has a 99.9% availability SLO and a review window of 30 days. Its total error budget is therefore 0.1% of eligible requests. During the current period, the service has processed 2,000,000 eligible requests and 1,200 were unsuccessful.
The observed bad-request rate is 1,200 ÷ 2,000,000 = 0.06%. The service has consumed 60% of its 0.1% budget and has 40% remaining. That result alone does not approve a release. The team should check whether the failures are concentrated in a critical endpoint, whether the rate is increasing, and whether the proposed change affects the same dependency.
If the release is a small, reversible configuration change with a staged rollout and a tested rollback, the team may proceed with close monitoring. If it changes the failure-prone dependency, a pause or additional mitigation is more defensible even though budget remains.
Example 2: Burn rate during an incident
Assume the same service normally consumes its budget evenly over 30 days. A short outage causes a sharp increase in failed requests over 20 minutes. The short-window burn rate is well above 1, while the monthly view is still below the total budget limit.
The two views answer different questions. The monthly view indicates how much budget remains. The short window indicates that the current condition is unsafe to ignore. The release decision should therefore focus first on containment: stop a related rollout if appropriate, restore service, validate recovery, and identify whether the incident is still consuming budget. Remaining monthly budget should not be used to justify continuing a known failure mode.
SLO review and error-budget policy template
Copy this policy into the team’s service documentation and adapt the thresholds:
Scope: This policy applies to [service] and the SLOs listed in [dashboard or catalog].
Review cadence: The team reviews SLO performance [weekly or monthly] and after every material incident or launch.
Release rule: Releases may proceed when the SLO is healthy, the change has an owner, and rollback or mitigation steps are tested.
Elevated consumption: When burn rate or customer impact exceeds the team’s agreed threshold, the release owner must reduce exposure, add safeguards, or obtain an explicit exception.
Budget exhausted: The team prioritizes restoration and reliability work before discretionary changes, except for changes needed to recover the service or address urgent risk.
Exceptions: An exception records the decision maker, customer impact, reason, duration, mitigations, and follow-up date.
Reset: SLO targets, SLIs, exclusions, and thresholds are reviewed after a planning cycle or major product change.
Monthly review agenda
- Confirm that the SLI query, denominator, exclusions, and dashboard still match the service.
- Review SLO attainment, budget consumption, and short- and long-window burn rates.
- Compare incidents, deployments, infrastructure changes, and support signals.
- Separate customer impact from internal symptoms and identify gaps in instrumentation.
- Decide which releases proceed, require additional safeguards, or are paused.
- Assign recovery actions with owners, due dates, and a verification measure.
- Record whether the SLO remains appropriate for the current product and platform design.
When to recalculate
Recalculate the worksheet whenever the underlying inputs change, not only at the scheduled review. Revisit it after a major incident, a new customer journey, a traffic or tenancy model change, a regional expansion, a dependency replacement, or a significant Kubernetes or cloud-platform change. Recheck it when observability queries, sampling, alert rules, aggregation, or exclusion logic changes.
Use the same discipline for release tooling. A new deployment strategy, image-tagging convention, rollback mechanism, or CI/CD gate can change the risk profile even if the SLO target stays fixed. The Docker image tagging strategy guide and OpenAPI validation workflow provide useful companion checks for release traceability and interface risk.
Before the next decision, assign one person to update the inputs and another to challenge the assumptions. Then record a clear outcome: release, release with safeguards, pause and recover, or change the SLO after review. A concise decision log makes the error budget useful to engineering, platform, product, and incident-response teams instead of leaving reliability data in a dashboard no one consults.