Availability: free-recall review
Free-recall prompts across the availability unit. Answer each in your own words first — including the arithmetic — then reveal the model answer and compare: SLO/SLA, the nines and error budgets, redundancy and correlated failure, failover and split-brain.
Retrieval beats re-reading. For each prompt, say or write a full answer from memory — including the arithmetic — before you open the model answer. The effort of reconstructing the SLO math, the redundancy sizing, and why a failover causes split-brain is what makes them stick.
Reconstruct the unit’s spine without looking back: how the three service-level terms relate and where to set each, what the nines cost and how availability composes, what an error budget buys you, how to size redundancy and why same-AZ replicas aren’t redundant, and how a failover can cause the very outage it was meant to prevent.
- 01Define SLI, SLO, and SLA, state how they relate, and where you set the SLO.
- 02Give the nines table for downtime per month and explain why each nine costs ~10× more.
- 03What is an error budget and how does it change a team's behavior?
- 04Explain redundancy sizing (N/N+1/N+2), active-active vs active-passive, and why same-AZ replicas aren't redundant.
- 05Distinguish failover from fault tolerance, explain split-brain and its fix, and the retry-storm trap.
If you could reconstruct each answer from memory, you hold the unit’s spine. The three service-level terms relate one way — measure the SLI, target the SLO, promise the SLA — and the SLO sits tighter than the SLA as an early warning. The nines (99.9% ≈ 43 min/month → 99.999% ≈ 26 s) cost an order of magnitude each because availability multiplies across dependencies, and the error budget (100% − SLO) turns reliability into a spendable resource. Redundancy is sized to N+1 against peak and placed in independent failure domains, because same-AZ replicas are correlated failure, not redundancy. Failover (reactive, with a recovery window) differs from fault tolerance (masked, no window), and a failover that can’t tell dead from unreachable causes split-brain — fixed by quorum or fencing, not better timeouts — while the retry storm at failover is tamed by bounded timeouts and capped retries with backoff and jitter. The unifying idea: availability ≈ MTBF/(MTBF+MTTR), so since failures are inevitable, the better investment is usually fast, rehearsed recovery — and all of it is reasoning you do before the incident, not behavior you discover during it.
Something unclear?
Ask a question about this lesson. Questions are anonymous and go straight to the author to make the lesson better.