Aleksey Charapko revisits metastable failures, defining them as self-sustaining performance issues caused by positive feedback loops. A classic example is a retry storm, where client retries amplify server overload, creating a vicious cycle. The core problem lies in ambiguous signals—timeouts can indicate either transient faults or overload, leading clients to take the wrong action.

The article breaks down system interactions into actions, states, and signals. When a serving system is overloaded, it emits timeout signals. Clients misinterpret these signals and retry, worsening the load. This creates a control loop where ambiguous signals drive erroneous actions, making recovery difficult. Avoiding these loops entirely is often impossible in complex, economical systems.

Practical mitigation strategies include minimizing unnecessary interactions and avoiding actions that create positive feedback. However, some actions, like retries for transient faults, are semantically necessary. The key is reducing signal ambiguity—using multiple signals to trigger actions. While complete avoidance may be impractical, these strategies help contain metastable failures in production distributed systems.