1. HA vs DR

• High Availability (HA) — Minimize downtime during normal operations. Goal: near-zero interruption (seconds to minutes).
• Disaster Recovery (DR) — Recover from catastrophic failure (region outage, data loss). Goal: defined RTO/RPO.
• HA handles component failures transparently; DR handles site-level failures with explicit failover.
• Both are needed — HA doesn't protect against regional disasters; DR alone doesn't prevent everyday outages.

2. SLA Composition Math

Serial dependencies — Composite SLA = SLA_A × SLA_B × SLA_C. Every component in the chain reduces overall availability.

• Key insight: 3 services at 99.9% each → composite = 99.7% (≈ 26 hrs downtime/year, not 8.7 hrs).
• Parallel redundancy improves SLA: P(both fail) = (1 − SLA)². Two 99.9% instances → 99.9999%.
• Always calculate composite SLA for critical paths — the exam tests this arithmetic.

3. Patterns to Increase Availability

• Redundancy — Deploy multiple instances behind a load balancer.
• Active-Active — All instances serve traffic simultaneously; highest availability.
• Active-Passive — Standby instance activated on failure; simpler but higher RTO.
• Health probes — Load balancers and Traffic Manager detect unhealthy instances and reroute.
• Availability Zones — Spread across fault-isolated datacenters within a region.
• Queue-based load leveling — Decouple producers/consumers to absorb spikes.

4. Azure SLA Tiers

• Single VM (Premium SSD) — 99.9% (≈ 8.7 hrs/year downtime)
• Availability Set — 99.95% (≈ 4.4 hrs/year)
• Availability Zones — 99.99% (≈ 52 min/year)
• Multi-region active-active — Target 99.999% (≈ 5 min/year)

Each tier adds cost and complexity — design for the SLA the business actually needs.

5. Multi-Region Active-Active

• Front Door routes traffic to the nearest healthy region based on latency.
• Data layer: Cosmos DB multi-region writes or SQL Always On AG with readable secondaries.
• Challenge: Conflict resolution for concurrent writes; eventual consistency trade-offs.
• Cost: Double compute + premium data replication — justified only for mission-critical workloads.

6. Multi-Region Active-Passive

• Use Azure paired regions for sequenced updates and priority recovery.
• Automate failover with Traffic Manager health probes + Azure Automation runbooks.
• Test failover regularly — an untested DR plan is not a plan.

7. Health Modeling

• Composite health endpoint — A single /health API that checks all critical dependencies (DB, cache, messaging).
• Dependency mapping — Classify each dependency as critical (blocks requests) or degraded (graceful fallback).
• Return Healthy, Degraded, or Unhealthy — load balancers act on the response.
• Health probes should be lightweight — avoid heavy queries; cache dependency status briefly.
• Use Application Insights availability tests for external monitoring.

8. Chaos Engineering

• Principle: Inject controlled failures in production/staging to validate HA assumptions.
• Azure Chaos Studio — Managed service for running chaos experiments (VM shutdown, network faults, DNS failures).
• Define steady-state hypothesis → inject fault → observe → improve.
• Start in staging, graduate to production with blast-radius controls (target specific resource groups, abort conditions).
• Common experiments: kill an AZ, saturate CPU, add network latency, revoke DB credentials.

🧪 Knowledge Check