Introduction
High availability (HA) in data center design requires systematic thinking about failure modes, redundancy, and recovery. This guide covers architectural patterns for supremacy-enterprise-computing" title="Understanding Quantum Supremacy: What It Means for Enterprise Computing" class="internal-link">enterprise-grade availability.
Availability Targets
Understanding availability levels:
| Target | Downtime/Year | Use Case |
|---|---|---|
| 99% | 3.65 days | Development |
| 99.9% | 8.76 hours | Business apps |
| 99.99% | 52.56 minutes | Critical systems |
| 99.999% | 5.26 minutes | Financial/Healthcare |
Each additional "9" requires exponentially more investment.
Failure Domain Analysis
Identify Failure Domains
Map GPU Clusters vs Cloud Services" class="internal-link">infrastructure to failure domains:
- Server: Single machine failure
- Rack: Power/network to rack
- Row: Cooling or power distribution
- Room: Fire suppression, cooling
- Building: Site-wide events
- Region: Geographic disasters
Design for Failure Boundaries
Place redundant components across failure domains:
- Replicas on different racks
- Services across multiple rooms
- Data across geographic regions
Power Infrastructure
N+1 Redundancy
Minimum for production systems:
- N generators where N-1 can handle full load
- Redundant UPS systems
- Automatic transfer switches
2N Redundancy
For critical systems:
- Fully independent power paths
- Separate utility feeds
- Complete generator redundancy
Power Distribution
Utility A ─┬─ ATS ─┬─ UPS A ─┬─ PDU A ─ Rack A
│ │ │
Utility B ─┴─ ATS ─┴─ UPS B ─┴─ PDU B ─ Rack A
Network Architecture
Spine-Leaf Topology
Modern data center networks use spine-leaf:
- Spine switches: High-capacity core
- Leaf switches: Connect servers
- Every leaf connects to every spine: Provides path redundancy
Network Redundancy
- Dual-homed servers: Two NICs to different leaves
- MLAG/vPC: Active-active switch pairs
- Multiple ISP connections: Internet redundancy
- BGP failover: Automated route management
Storage Resilience
RAID and Beyond
RAID protects against disk failure but not controller/server failure:
- RAID 10: monitoring-practical-introduction" title="eBPF for Performance Monitoring: A Practical Introduction" class="internal-link">Performance and redundancy
- Erasure coding: Distributed redundancy
- Replication: Cross-site data copies
Storage Architecture Patterns
Pattern 1: Shared Storage
- SAN/NAS with redundant controllers
- Suitable for traditional workloads
- Limited scale
Pattern 2: Distributed Storage
- Software-defined storage (Ceph, MinIO)
- Horizontal scaling
- Higher complexity
Pattern 3: Hyperconverged
- Compute and storage combined
- Simplified management
- Vendor lock-in risk
Server-Level HA
Hardware Redundancy
Modern servers include:
- Redundant power supplies
- Hot-swappable drives
- ECC memory
- Multiple NICs
Clustering
Application clustering patterns:
- Active-Passive: Standby takes over on failure
- Active-Active: Load distributed across nodes
- N+M: M spare nodes for N active nodes
Monitoring and Recovery
Detection
- Hardware health monitoring (IPMI/BMC)
- Application health checks
- Network monitoring
- Log analysis
Automated Recovery
- Service restart on failure
- Failover to standby
- Auto-scaling to replace failed instances
- Self-healing infrastructure
Testing HA Systems
Chaos Engineering
Regularly test failure scenarios:
- Kill random processes
- Introduce network latency
- Fail storage components
- Simulate power loss
Disaster Recovery Drills
- Scheduled DR tests
- Documented procedures
- Measured RTO/RPO
Cost Considerations
HA investment increases non-linearly:
- 99% to 99.9%: 3-5x cost
- 99.9% to 99.99%: 5-10x cost
- 99.99% to 99.999%: 10x+ cost
Match investment to business requirements.
Conclusion
High availability requires holistic design across power, network, storage, and compute layers. Start with clear availability targets, map failure domains, and implement appropriate redundancy. Regular testing validates that designs meet objectives.