Monitoring and drift

error budget = 1 − SLO; burn rate = observed error rate ÷ (1 − SLO); time to exhaust = window ÷ burn rate

From the Google SRE workbook: burn rate 1 spends the budget exactly over the window; higher spends it sooner. The workbook's starting alert set for a 99.9% SLO pages at burn rate 14.4 over one hour (2% of budget) and 6 over six hours (5%), and opens a ticket at burn rate 1 over three days (10%).

Worked example: SLO 92% pass rate on a 30-day window: budget 8%. A week with 14% failures burns at 1.75, so the budget would be gone after about 17 days.

PSI = Σ over bins (a_i − e_i) × ln(a_i ÷ e_i) = KL(a ‖ e) + KL(e ‖ a)

a_i is the share of this week's scores in bin i, e_i the share at launch. It is the symmetric sum of the two Kullback-Leibler divergences, KL(p ‖ q) = Σ p_i ln(p_i ÷ q_i). Rules of thumb read below 0.10 as little shift and 0.10 to 0.25 as a significant shift requiring action, though those cut-offs were set without reference to error rates.

Worked example: Launch shares (0.05, 0.10, 0.20, 0.35, 0.30) against week 6 shares (0.12, 0.18, 0.25, 0.28, 0.17): PSI ≈ 0.21.

n ≈ z² × p × (1 − p) ÷ E², with z = 1.96 for a 95% interval

How many sampled conversations a weekly review needs to estimate a pass rate p to within E points. It follows from the standard error of a proportion; see the error-bars topic.

Worked example: p = 0.9, E = 0.03: 3.84 × 0.09 ÷ 0.0009 ≈ 384 cases a week. Accept ± 5 points and about 138 will do.

budget consumed by an alert window = burn rate × window ÷ period

The same relation, rearranged, says how much of the budget a burst of errors has already spent by the time an alert fires.

Worked example: Burn rate 36 over one hour of a 30-day period: 36 × 1 ÷ 720 = 5% of the budget, the workbook's own example.