When trying again makes things worse.
WHAT RETRIES SOLVE
Transient failures.
WHAT THEY CAUSE
Duplicates, if operations are not idempotent Amplified load on a struggling service Cost, on metered interfaces
WHAT A RETRY STORM IS
Many clients retrying simultaneously against a recovering service.
WHAT IT DOES
Prevents recovery.
WHAT PREVENTS IT
Increasing delays Random variation in those delays A limit on attempts A circuit breaker
WHY RANDOM VARIATION MATTERS
Without it, clients synchronise and retry together.
WHAT TO NEVER RETRY
Requests rejected as invalid.
WHY
They will be rejected identically.
WHAT TO BE CAREFUL ABOUT RETRYING
Anything creating or charging.
WHAT MAKES IT SAFE
An idempotency key.
WHAT TO CONSIDER ABOUT NESTED RETRIES
Retries at several layers multiply.
WHAT THAT PRODUCES
Three attempts at each of three layers becoming twenty-seven.
WHAT TO DO
Retry at one layer, deliberately chosen.
WHAT TO LOG
Every retry, so the pattern is visible.
WHAT TO MONITOR
Retry rate.
WHY
A rising rate indicates a developing problem before failures appear.
