Availability
Nines and downtime budgets, removing single points of failure, redundancy patterns across zones and regions, and how dependencies multiply unavailability.
What it is
Availability is the proportion of time (or of requests) for which a system does its job. Modern practice measures it per request: successful requests ÷ valid requests over a window, against an SLO.
| Availability | Downtime per year | Per month (30 days) |
|---|---|---|
| 99% | 3.65 days | 7.2 hours |
| 99.9% | 8.76 hours | 43.2 minutes |
| 99.95% | 4.38 hours | 21.6 minutes |
| 99.99% | 52.6 minutes | 4.3 minutes |
| 99.999% | 5.26 minutes | 26 seconds |
The dependency math
If a request needs services A, B and C in series, each 99.9% available and failing independently, end-to-end availability is 0.999³ ≈ 99.7%. Serial dependencies multiply unavailability. Redundant parallel components improve it: two independent 99% replicas give 1 − 0.01² = 99.99% (if failures are truly independent, which correlated failures often break).
flowchart LR
subgraph Serial["Serial: 99.9% × 99.9% × 99.9% ≈ 99.7%"]
A1[Service A] --> B1[Service B] --> C1[Service C]
end
subgraph Parallel["Parallel replicas: 1 − 0.01² = 99.99%"]
LB[Load balancer] --> R1[Replica 99%]
LB --> R2[Replica 99%]
endTechniques
- Redundancy at every tier: multiple instances, multiple availability zones (AZs), replicated databases with automated failover.
- Health checks and automatic failover: load balancers remove unhealthy instances; database managed failover (e.g. Aurora, Cloud SQL HA, Azure SQL zone redundancy).
- Remove synchronous dependencies from critical paths: make non-essential calls asynchronous or optional (graceful degradation).
- Static stability: the system keeps working when the control plane is impaired (pre-provisioned capacity, cached config). AWS uses this term in its builders' library.
- Safe deployments: canary, blue/green, automatic rollback. Most outages are caused by changes.
- Cell-based architecture to limit blast radius.
Real-world example
A payments API targets 99.99%. It runs active-active across three AZs with at least 50% spare capacity per AZ so that losing one AZ needs no scaling. The fraud-scoring dependency has a 50 ms timeout and a fallback to a rules-only score, so a fraud-service outage degrades accuracy, not availability. Deployments go to one cell at a time with automated rollback on error-rate increase.
Advantages and disadvantages of higher availability
| Advantages | Disadvantages |
|---|---|
| Revenue and trust protected | Cost grows sharply per extra nine |
| Meets regulatory/contractual needs | Multi-region adds consistency and complexity issues |
| Better user experience | More components to operate and test |
Technology options
Multi-AZ load balancers (ALB/NLB, Azure Load Balancer/Application Gateway, Google Cloud Load Balancing), managed HA databases, global traffic managers (Route 53, Azure Front Door/Traffic Manager, Google global LB), Kubernetes with pod anti-affinity and PodDisruptionBudgets, chaos tools (AWS FIS, Azure Chaos Studio, Gremlin, Chaos Mesh).
Architect-level considerations
[!ARCHITECT] Define availability per user journey, not per system. "Browse" and "checkout" deserve different targets. Draw the dependency chain for each critical journey and compute its theoretical availability.
[!EA] Publish platform availability tiers and require each system to declare its tier, dependencies and DR pattern in the architecture review.
[!CTO] Ask: what does one hour of downtime cost? That number justifies (or not) active-active multi-region.
Interview questions and answers
How would you take a system from 99.9% to 99.99%?
Find the top causes of past downtime (usually deployments, dependencies, database failover time, config errors). Add multi-AZ redundancy everywhere, faster failover, canary deploys with automatic rollback, timeouts and fallbacks for dependencies, and SLO-based alerting. Only then consider multi-region.
Is a system with 99.99% SLA dependencies able to offer 99.99%?
Not if it depends serially on several of them. Either reduce critical-path dependencies, add redundancy across providers or regions, or degrade gracefully when they fail.
Availability vs reliability?
Availability is whether it is up and answering; reliability is whether it does the right thing correctly over time (including not losing or corrupting data). A system can be available but unreliable, for example returning wrong balances.