tenacity-python

Use the @retry decorator with a narrow predicate like retry_if_exception_type, a finite stop such as stop_after_attempt, and a wait strategy like wait_random_exponential. Set reraise=True so callers receive the final domain exception instead of RetryError.

Decorate an async def function with @retry or use AsyncRetrying for block iteration. The awaited client call and sleep must remain nonblocking; never call time.sleep inside an async attempt because Tenacity cannot make blocking code asynchronous.

Do not retry permanent failures such as authentication errors, invalid input, or programmer defects, and never retry non-idempotent writes without an idempotency key or transactional upsert. Also avoid retrying broad Exception, cancellation, or keyboard interrupts.

Inject sleep=lambda _: None into Retrying or replace the wait strategy only in the test-owned policy. Feed deterministic exception sequences and assert attempt counts, terminal exceptions, and that visible side effects occur at most once.

No, stop_after_attempt(3) permits at most three total attempts including the first call, not three retries after it. The first invocation counts as attempt 1 in Tenacity's retry state model.

Duplicates occur when a write lacks an idempotency mechanism or when nested layers each retry, multiplying attempts. Require a stable idempotency key or atomic upsert for writes, and designate a single retry owner with one outer deadline.