map / distributed-systems / availability-math

Distributed Systems · principle 2 of 6

Availability math

Dependencies in series multiply their availabilities down; redundant copies in parallel multiply their failure probabilities down.

01 · Why it’s true

If a request needs several components to all work, it succeeds only if every one of them succeeds. When failures are independent, you multiply their availabilities — and the product is always lower than the weakest link. Five components at 99.9% each give about 99.5%: roughly 3.6 hours of downtime per 30 days instead of 43 minutes.

Redundancy works the other way. If any one of several independent copies is enough, the request fails only if all of them fail at once. Two copies that are each 99% available give 1 − 0.01 × 0.01 = 99.99%. The catch is that copies often share failure causes (same deploy, same region, same config), so real-world gains are smaller.

02 · Picture it

1Load balancer99.99%×2App99.9%×3Database99.95%=4End to end≈ 99.84%In series, availabilities multiply: 0.9999 × 0.999 × 0.9995 ≈ 0.9984 — lower than anysingle part. In parallel, failure chances multiply instead: two 99% copies give 99.99%.
1Load balancer99.99%×2App99.9%×3Database99.95%=4End to end≈ 99.84%In series, availabilities multiply:0.9999 × 0.999 × 0.9995 ≈ 0.9984 —lower than any single part. Inparallel, failure chances multiplyinstead: two 99% copies give 99.99%.
A request path is only as available as the product of its parts.

03 · In the interview

  • Setting realistic SLOs for a service that depends on others.
  • Turning an SLO into an error budget: 99.9% over 30 days is about 43 minutes of failure.
  • Explaining why redundancy only helps when failures are truly independent.

04 · Practice question

“Our service calls five internal dependencies on every request, each with a 99.9% SLO. What SLO can we promise?”

A strong answer covers

  1. In series, multiply: 0.999⁵ ≈ 0.995, or about 99.5% — around 3.6 hours of failure per 30 days.
  2. So promising 99.9% end to end isn’t realistic without changing the design.
  3. Options: remove dependencies from the critical path, cache or degrade gracefully when one fails, make calls asynchronous, or add redundant paths.
  4. Mention the caveat: shared causes (a bad deploy, a network event) make failures correlated, not independent.

Want to play with it? Open the interactive deep dive →