OSLTC.008 teaches technical leaders how to choose reliability targets and use error budgets to decide when to release features, improve systems or escalate service risk. Site reliability engineering (SRE) replaces vague promises such as “never go down” with measurable objectives tied to user experience. Three related concepts matter: service-level indicators (SLIs), service-level objectives (SLOs) and service-level agreements (SLAs). The framework is useful for web services, internal APIs and production platforms, but its measurements must match what users actually experience.
Earlier lessons covered incident leadership and blameless postmortems. An SLO adds an explicit definition of acceptable service behavior before the next incident occurs. Google’s SRE Workbook recommends starting with indicators relevant to the user rather than every available infrastructure metric.
SLI, SLO and SLA are not interchangeable
An SLI measures behavior, such as the fraction of eligible API requests served successfully. An SLO sets the target: for example, at least 99.9% of eligible requests should succeed over a rolling 30-day period. An SLA is a contractual agreement that may specify remedies when a service fails its commitments. Google’s comparison separates the three concepts.
The numerator and denominator are critical. Monitoring must define which requests count, which outcomes qualify as good, and how missing telemetry is handled. A CPU-utilization graph is not automatically a valid availability SLI. A user-facing latency objective can also be valuable even when availability is high.
Worked example: a 99.9% error budget
Imagine a fictional service receiving 1,000,000 eligible requests in 30 days. A 99.9% success objective permits at most 1,000 failed requests: 1,000,000 × (1 − 0.999) = 1,000. After 300 failures, the service has used 30% of its allowed failures and has 700 remaining. Those values assume a request-based SLI and fixed observed request volume, not a guaranteed number of outage minutes.
Changing the target to 99.99% would permit only 100 failures for the same request count. Reliability targets therefore represent consequential engineering and staffing decisions. A more stringent SLO is not automatically better if customers have no corresponding need or the costs crowd out valuable work. Google’s example error-budget policy connects an exhausted budget to a documented response rather than blaming individuals.
Burn rate measures how quickly risk accumulates
Burn rate = observed failure fraction ÷ allowed failure fraction. With a 99.9% SLO, the allowed fraction is 0.1%. An observed failure fraction of 0.6% produces a 6× burn rate. Google’s SLO alerting chapter gives reference examples such as 14.4× over a one-hour window and 6× over a six-hour window, commonly paired with shorter confirmation windows. Alert thresholds depend on traffic, time horizon and operational priorities; they are not universal paging rules.
Turn measurements into responsible decisions
A leader should establish the affected customer journey, objective period, owner, alerting approach and written response policy before using an error budget to influence a release. Rapid burn may justify pausing a risky rollout; healthy margin may permit controlled experimentation. The decision should consider severity and customer impact, not simply a dashboard color. OSLTC.005 on accountability and OSLTC.002 on expectations explain how to assign ownership without confusing accountability with punishment.
For a read-only exercise, calculate an error budget from a sample of successful and failed requests, chart budget use across four weeks, and compare whether short and long alert windows would trigger. State assumptions explicitly. Learning conclusion: reliable technical leadership turns service measurements into transparent trade-offs while preserving evidence, customer focus and a blameless improvement culture.

Leave a Reply