A metastable failure is a state of persistent overload with very low goodput that continues after the original trigger has been removed, because a sustaining feedback loop inside the system keeps it overloaded.
Example
A brief traffic spike makes requests exceed client timeouts. Clients retry, so offered load stays above capacity even after demand returns to normal. Restarting servers or adding retries does not help; shedding load and limiting retries does.
See Metastable Failures: When Recovery Itself Keeps the System Down.