SLO Burn-Rate Alerts: Practical Thresholds, PromQL Patterns, and Alert Review Checklist
SREObservabilitySLOsPrometheusAlerting

SLO Burn-Rate Alerts: Practical Thresholds, PromQL Patterns, and Alert Review Checklist

CChallenges.pro Editorial Team
2026-11-13
7 min read

Learn how to calculate SLO burn rates, choose multi-window thresholds, write PromQL patterns, and review alert quality without adding paging noise.

SLO burn-rate alerts help teams page on sustained reliability risk rather than every isolated error. This guide explains how to estimate error-budget consumption, choose fast- and slow-burn windows, write adaptable PromQL alert patterns, and review alert quality before changing thresholds or routing.

Overview

An error budget is the amount of unreliability permitted by an SLO during a defined period. For an availability SLO, a 99.9% target over 30 days permits 0.1% of requests to be unsuccessful. A burn rate describes how quickly the service is consuming that allowance compared with the planned rate.

The basic calculation is:

Burn rate = observed bad-event ratio ÷ allowed bad-event ratio

If the observed error ratio is 1% and the SLO allows 0.1%, the service is burning its budget at 10 times the normal rate. A burn rate of 1 means the entire budget would be consumed at the intended rate by the end of the SLO period. A burn rate above 1 indicates faster-than-planned consumption.

One threshold rarely works for every failure mode. A short window detects severe incidents quickly, while a longer window confirms that the problem is persistent enough to justify interrupting an engineer. Combining both windows is the central idea behind multi-window burn-rate alerting:

  • Fast-burn alerts: catch incidents that could exhaust a large portion of the budget quickly.
  • Slow-burn alerts: catch smaller but sustained regressions that may be missed by short windows.
  • Severity separation: route urgent, high-confidence risk to paging and less urgent risk to tickets or team channels.

Thresholds are control settings, not universal constants. They should reflect your SLO period, traffic volume, response expectations, and tolerance for missed risk versus unnecessary pages.

How to estimate

1. Calculate the allowed error ratio

Convert the SLO target into an allowed bad-event ratio:

Allowed ratio = 1 − SLO target

For a 99.9% SLO, the allowed ratio is 0.001. Use the same event definition in the numerator and denominator. For example, count failed requests against all eligible requests, rather than comparing failed requests with a different traffic metric.

2. Measure the observed ratio

For a request-based SLO, calculate:

Observed ratio = bad requests ÷ total eligible requests

For time-slice or availability models, use the corresponding bad-time and total-time measurements. The alert is only as reliable as the SLI behind it: retries, redirects, excluded endpoints, synthetic checks, and health probes must be handled consistently.

3. Convert burn rate into a budget horizon

A useful planning estimate is:

Approximate time to consume the full budget = SLO period ÷ burn rate

With a 30-day SLO period and a burn rate of 15, the entire budget would be consumed in roughly two days if the rate continued. This is an estimate, not a prediction. Traffic changes, recovery actions, and intermittent failures can make the actual outcome different.

Reference thresholds

Use caseExample long windowExample short windowTypical response
Fast burn1 hour5 minutesPage when both windows indicate severe risk
Moderate burn6 hours30 minutesPage or urgent team notification, depending on policy
Slow burn24 hours2 hoursCreate a ticket or notify the service channel

These windows are starting points for review, not a prescribed standard. A team that needs rapid intervention may choose shorter windows. A low-volume service may need longer windows to avoid decisions based on a handful of requests.

Inputs and assumptions

Before writing a rule, document the inputs that determine its behavior:

  • SLO target and period: record the target, such as 99.9%, and the evaluation period used by the team.
  • SLI definition: specify which requests, jobs, messages, or time intervals count as eligible and bad.
  • Traffic floor: decide whether a minimum request or event rate is required before paging. Ratios from very little traffic can be unstable.
  • Telemetry completeness: determine what happens when metrics stop arriving. Missing data should not silently appear as perfect availability.
  • Evaluation interval: choose a rule evaluation frequency that is appropriate for the shortest alert window.
  • Routing and ownership: identify the service owner, escalation path, and destination for pages, tickets, and informational notifications.
  • Recovery expectations: connect severity to the time available before meaningful budget damage occurs.

A generic PromQL pattern for a request-based SLI can be written as follows. Metric names and labels must be replaced with those used by your instrumentation:

sum(rate(http_requests_total{service="checkout",status_class="5xx"}[5m]))
/
sum(rate(http_requests_total{service="checkout"}[5m]))
> (1 - 0.999) * 14.4

This compares the recent bad-request ratio with the allowed ratio multiplied by an example burn-rate threshold. In practice, calculate the ratio over both a long and a short window and require both conditions:

(bad_ratio_1h > allowed_ratio * fast_burn)
and
(bad_ratio_5m > allowed_ratio * fast_burn)

Prometheus alert rules generally use recording rules to calculate reusable ratios and alert expressions to combine windows. This improves readability and makes it easier to inspect each component during an incident. Confirm label matching, counter resets, aggregation, and missing-series behavior in a non-paging environment before enabling notifications.

Worked examples

Example 1: Severe, short-lived failure

Assume a 99.9% availability SLO, an allowed ratio of 0.001, and a fast-burn threshold of 14.4. The threshold ratio is 0.0144, or 1.44%. If both the five-minute and one-hour bad-request ratios exceed 1.44%, the service is showing a concentrated failure large enough to justify a page under this policy.

The two-window requirement reduces the chance that a single scrape, deployment event, or tiny traffic sample creates an incident. It does not eliminate false positives, so inspect the traffic floor and telemetry quality before treating the threshold as final.

Example 2: Persistent degradation

Suppose the same SLO experiences a 0.2% bad-request ratio for several hours. That is twice the allowed ratio, so the burn rate is 2. It may not trigger a fast-burn page, but it steadily consumes budget. A six-hour and 30-minute pair, or a 24-hour and two-hour pair, can surface this condition for investigation before the monthly budget is materially reduced.

In this case, a ticket with a runbook link may be more appropriate than a page if the estimated time to exhaust the budget exceeds the team's response policy. The choice should be explicit rather than an accidental result of whichever rule was easiest to write.

Example 3: Low-volume service

A service receiving only a small number of requests may produce a 100% error ratio from one failed request. A ratio-only alert can therefore be noisy or misleading. Consider a minimum traffic condition, longer windows, or an SLI that better represents meaningful user impact. Record the trade-off: suppressing low-volume alerts may delay detection, while paging on every isolated event may train engineers to ignore the alert.

When to recalculate

Review burn-rate alerts at every SLO review and after any material change to reliability policy, traffic, architecture, or paging. Recalculate when:

  • the SLO target or evaluation period changes;
  • traffic volume, seasonality, or request distribution changes;
  • an endpoint, dependency, retry policy, or status classification changes;
  • telemetry is migrated, labels are renamed, or metric cardinality changes;
  • alert fatigue investigations show repeated false positives or missed incidents;
  • ownership, escalation rules, on-call coverage, or runbook locations change.

Use this review checklist before enabling or revising a rule:

  1. Can the team state exactly what counts as a bad and eligible event?
  2. Are the long and short windows appropriate for the service's traffic and recovery time?
  3. Does the rule handle missing telemetry distinctly from healthy telemetry?
  4. Is there a minimum traffic safeguard where ratios could be unstable?
  5. Are both windows required for the intended severity?
  6. Does the alert route to a current owner with a clear escalation path?
  7. Does the notification include the SLO, observed ratio, burn rate, affected dimensions, and time windows?
  8. Does it link to an incident response runbook with first checks and rollback or mitigation steps?
  9. Has the rule been tested against historical incidents and quiet periods?
  10. Is the resulting page actionable for the person receiving it?

Keep the calculation, PromQL recording rules, routing configuration, and runbook together in version control. After each review, compare alert outcomes with actual incidents and error-budget consumption. Adjust one assumption at a time where possible, then document why the change improves detection or reduces unnecessary interruption. For broader reliability planning, pair these alerts with an SLO and error budget template and use an incident troubleshooting decision tree when delivery changes are a suspected cause.

Related Topics

#SRE#Observability#SLOs#Prometheus#Alerting
C

Challenges.pro Editorial Team

DevOps and SRE Editors

Senior editor and content strategist. Writing about technology, design, and the future of digital media. Follow along for deep dives into the industry's moving parts.