With exponential backoff, each failed attempt waits longer than the previous one before retrying. Adding jitter prevents many clients from retrying at exactly the same moment and recreating the overload.
Backoff must be bounded by maximum attempts, total time and cost. It is appropriate only for errors likely to be temporary; validation or authorization failures should normally stop immediately rather than enter a retry loop.
ELI5
Exponential backoff retries a temporary failure with longer waits after each attempt. The growing delay reduces pressure on a service that may already be overloaded.
For example, a client can wait about one second after the first failure, two seconds after the next and four seconds after another, with small random changes so many clients do not retry together. It should stop after a clear limit and should not retry invalid requests.
