Availability and Reliability

A service can respond and still behave incorrectly

Availability and reliability are often mentioned together, but they are not the same thing. A system can be available — accepting requests and returning responses — while still producing incorrect data. And the opposite can happen: the system may preserve data correctly but be temporarily unavailable to the user.

Availability: can the system be used right now?

Availability shows what fraction of time a system is ready to serve requests. If a service works almost all the time and recovers quickly from rare failures, its availability will be high.

availability = uptime / total time

Availability is affected by more than complete application crashes. If a service is technically running but every request times out, from the user's perspective it is effectively unavailable as well.

Reliability: can the result be trusted?

Reliability describes a system's ability to perform its work correctly and consistently over time. For different products, “correctly” means different things: not losing a payment, not delivering the same event twice, preserving an uploaded file, or returning the correct order state.

Reliability becomes especially visible not during normal operation, but when problems occur: a process restart, temporary network loss, a dependent service failure, or hardware malfunction.

Failures should be treated as a normal part of the system

The more components there are, the greater the chance that one of them will misbehave at a particular moment. That is why it is useful to design a distributed system not with the thought “servers do not fail,” but with the thought “individual failures are inevitable — what happens after one occurs?”

Client → Service A → Service B → Database
                    X
             Service B unavailable

The architecture must define what to do in such a situation: can the operation be retried, should a temporary error be returned, can a standby instance be used, or can the system continue in a degraded mode? The exact mechanisms depend on the system requirements.

An availability number always has context

The phrase “99.9% availability” is incomplete by itself. You need to understand the period over which it is calculated, what counts as a successful request, which components are included in the measurement, and what the end user actually experiences.