← All cheatsheets

Reliability & Operations

Availability & SLO Math

What each “nines” target costs in real downtime (99.9% = 8.8 h/year), how serial dependencies multiply to erode it, and why tail latency — not the average — is what your SLO should track.

Updated August 30, 2026 · 4 min read

Downtime per “nines”

99% (two nines)~3.65 days / year~7.2 h / month
99.9% (three nines)~8.76 h / year~43 min / month
99.99% (four nines)~52.6 min / year~4.3 min / month
99.999% (five nines)~5.26 min / year~26 s / month

How availability composes

Serial dependencies multiplyTwo 99.9% services in series → ~99.8%every hard dependency lowers the ceiling
Redundancy (parallel)n independent copies → 1 − (1 − a)ⁿhow you claw nines back
Error budget1 − SLO99.9% → 0.1% budget to spend on releases

Latency SLOs

Track percentiles, not the meanp50 / p95 / p99 / p99.9the mean hides the tail users feel
Fan-out amplifies the tailA request hitting 100 services sees each service’s p99one slow dependency dominates

In the interview

  • Convert the target to real downtime out loud: “four nines is ~52 minutes a year — that rules out manual failover.” The number drives the architecture.
  • Every synchronous dependency caps your availability at the product of theirs. Fewer hard dependencies, or make them async, to raise the ceiling.
  • Set SLOs on p99/p99.9, not averages. At scale the tail is the experience, and fan-out makes one slow service everyone’s problem.