Availability without Recovery: A Class of Failures Not Captured by SLAs

Abstract

Availability is usually regarded as the most important reliability goal: if the service stays "up,” then the users should be able to work, the revenue should continue, and their trust should be preserved. But being responsive to this single dimension, a system can still neglect the other crucial attribute of resilience recovery. Thus we highlight the gap between the real meaning of availability and the way it is referred to in the industry. While the SLAs only talk about system uptime or error rates, hardly ever, the capability of a system to quickly, safely, and fully recover from faults is taken into account. So that, even under a scenario of prolonged network degradation, recurring problems, or half-working devices that never seem to return to normal, the contract availability requirements can be considered met by the business. We define such a situation as availability without recovery: a scenario where services thus on paper avail themselves (fulfilling health checks and SLA criteria) but still manifest a constant incapacity to heal performance, correctness, or capacity post-disruptions. Besides setting up a framework for such a phenomenon, we give examples of the failure's fingerprints (e.g., repeatedly attempting an operation, backlog accumulating, reliance on services functioning at a degraded level, and "green dashboards" camouflage loss of functionality), and recovery-specific metrics complementary to conventional availability metrics. In a way through a distributed production system example from real life, we have demonstrated that systems which are inheriting dependency cascades, automations gone wrong, and suffering from a silent shortage of resources can keep high levels of availability at the same time that recovery cycles fail.