Course contentsShow
← System Design
Lesson 508 of 1,91914. CAP TheoremFree lesson

Availability is Also a Spectrum

Availability varies by operation; some requests may succeed while others fail, and latency impacts perceived availability.

Availability is Also a Spectrum

What you'll learn: How availability isn't simply "up" or "down"—different operations can have different availability levels, and latency affects whether users perceive your system as available.

Understanding Availability as a Range

Just as consistency exists on a spectrum, availability isn't binary either. Your system doesn't just flip between "100% available" and "completely down." Different parts of your system can exhibit different levels of availability simultaneously.

Operation-Level Availability

Not all requests succeed or fail uniformly. During a partial outage or partition, your system might:

  • Successfully serve read requests while rejecting writes
  • Handle cached data requests quickly but timeout on fresh queries
  • Process simple lookups but fail complex joins or aggregations
  • Serve critical features (login, checkout) while degrading secondary ones (recommendations, reviews)

This is selective availability—some operations remain available while others don't.

Latency as Hidden Unavailability

A system responding after 30 seconds might technically be "available," but users experience it as down. If your p99 latency spikes from 100ms to 10 seconds during a partition, you've lost practical availability even though responses eventually arrive.

Perceived availability = uptime + acceptable performance

The Real-World Picture

Consider an e-commerce site during a database partition:

  • Product browsing works (cached data): available
  • Search operates with stale results: degraded availability
  • Checkout times out after 10 seconds: functionally unavailable
  • Inventory updates fail: unavailable

Same system, same moment—four different availability states.

Key Takeaway: Availability operates on a spectrum across operations and includes latency; systems can be simultaneously available for some requests and unavailable for others, making "choose availability" a nuanced, operation-specific decision rather than a system-wide toggle.