Partial Availability and Graceful Degradation
What you'll learn: How to measure and account for partial outages where only some users or features are affected, rather than treating all failures as total system failures.
The Problem with Binary Availability
Traditional availability calculations assume a system is either fully up or completely down. But real-world failures are rarely so clean. What happens when only 20% of your users can't log in? Or when search works but recommendations fail? Treating these as "100% down" overstates the impact; ignoring them understates it.
Understanding Partial Availability
Partial availability recognizes that systems can be partially functional. If your e-commerce site serves 95% of requests successfully during an outage, that's not zero availability—it's 95% availability for that period.
To measure partial availability:
- Track successful requests vs total requests (not just up/down)
- Segment by user cohort, geography, or feature
- Weight by business impact (checkout failures matter more than color preference failures)
Graceful Degradation in Practice
Graceful degradation means designing systems to maintain core functionality when dependencies fail. Think of it like a car that loses air conditioning but still drives—degraded, not dead.
Examples:
- Disable personalization when the recommendation engine fails, but keep search working
- Show cached product listings when the inventory service is slow
- Accept orders without real-time fraud checks, flagging them for manual review later
This approach preserves partial availability rather than cascading into total failure.
Measuring What Matters
When calculating availability with partial outages:
- Weight by request volume: 10% of users failing for 1 hour ≠ 100% failing for 6 minutes
- Weight by user impact: premium users may count more than free users
- Consider feature criticality: checkout > reviews > recommendations
Key Takeaway: Partial availability and graceful degradation let you measure and maintain realistic system reliability by acknowledging that not all failures are total—designing for degraded states often delivers better user experience than binary up/down systems.