Service Level Objectives

The Hierarchy

 SLI (indicator): the MEASUREMENT
   "fraction of GET /feed requests serving < 300ms"

 SLO (objective): the TARGET on an SLI
   "99.5% of GET /feed requests < 300ms, over 28 days"

 SLA (agreement): the CONTRACT with consequences
   "if feed availability drops below 99.5%, credits issued"

 SLI ──target──► SLO ──penalties/business terms──► SLA
 you control:    SLIs and SLOs.
 lawyers control: SLAs. engineers should influence both.

Choosing Good SLOs

 principle 1: MEASURE WHAT USERS EXPERIENCE
   ✗ "CPU below 70%"          ← internal; users don't feel CPU
   ✗ "99.9% uptime"           ← up for whom? doing what?
   ✓ "99% of searches return results in <500ms"  ← felt quality

 principle 2: LEAVE ERROR MARGIN FOR INNOVATION
   target = user-happiness-threshold minus buffer:
   users happy at 99.9%? set SLO 99.5%.
   the gap is your error budget — fuel for velocity.

 principle 3: FEW AND SHARP
   3-5 SLOs per service max:
     availability (success rate), latency p99,
     freshness (data age), correctness/critical-check
   twenty vague SLOs govern nothing.

The Error Budget Connection

 SLO defines the budget:
   99.5% over 28d → 0.5% of requests may fail
   at 1M req/day → 50M × .005 = 250k failures/month allowed

 budget states drive behavior:

 budget remaining ██████░░ → ship features aggressively
 budget burning  ███░░░░░ → stabilize; risky changes pause
 budget spent    ░░░░░░░░ → reliability work ONLY;
                            releases freeze until recovered

 this converts reliability from vibes-vs-pressure arguments
 into a shared number both product and engineering obey.

Operating SLOs

Activity	Cadence
Burn-rate alerts (fast+slow windows)	real-time
Budget review in planning	per sprint
SLO review: still the right targets?	quarterly
SLA negotiation with legal/business	per contract cycle

 burn-rate alerting (multiwindow):
 fast burn: 2%/5min of budget → page NOW
 slow burn: 5%/6h              → ticket today
 avoids pager-fatigue of naive threshold alerts.

Interview Framing

“What are the SLOs for your design?” scored answer: 3–4 concrete user-facing SLIs with numbers, error-budget math shown, degradation behavior tied to budget state (“feed degrades to cached when budget half-spent”), distinction from SLA mentioned once. Designs with declared SLOs read as operated systems rather than science projects — declare them.