From SLI fractions to budget burn and alerts
Two fractions, one budget, two alert speeds
teaching thresholds: latency 99 percent under 300 ms, availability 99.9 percent, 30-day window
SLI fractions
latency: fast reads / good reads
availability: non-error / all
timeout = failure in both
dependency error hits availability
independent numerators:
a fast error is still an error
30-day error budget
teaching volume: 50M reads, budget 50,000
green: safe, amber: warn, red: spent
budget at zero means:
freeze releases, reliability only
burn alerts
fast-burn: page
2 percent of budget in 1 hour
action: rollback now
slow-burn: ticket
10 percent of budget in 6 hours
action: investigate this shift
SLO is the internal objective, SLA is the customer contract with remedies
example SLA: 99.5 percent monthly with 10 percent credit, deliberately looser than the SLO
legend: blue = measure, green = budget, peach = alert, yellow = contract split