Retries & backoff

parry retries transient failures with exponential backoff and jitter. It never retries errors that will fail again, such as an invalid key or a malformed request.

parry.install(max_attempts=3, retry_base_delay=0.5, retry_max_delay=30.0)
parry.install(retry=False)   # turn parry's retry layer off

Retry schedule

max_attemptsretry_base_delay (s)retry_max_delay (s)Retry-After header (s)

jitterRe-roll jitter

What gets retried

Retried	Never retried
RateLimitError (429)	BadRequestError (400)
APITimeoutError, APIConnectionError	AuthenticationError (401)
InternalServerError and every 5xx	PermissionDeniedError (403)
Anthropic overload (529)	NotFoundError (404)
408, 409, 425	UnprocessableEntityError (422)
httpx timeouts and connection errors	parry’s own BudgetExceeded / RateLimited

Errors are classified by HTTP status when there is one, and by exception class name otherwise. Retries exhausted? The SDK’s last exception is re-raised unchanged, so your existing except openai.RateLimitError: blocks still work.

The formula

For the n-th failure (counting from 1):

raw   = min(retry_max_delay, retry_base_delay × 2^(n-1))
delay = raw/2 + random(0, raw/2)        # "equal jitter": never an instant retry

If the response carries a retry-after-ms or retry-after header, that value is used instead, capped at retry_max_delay.

Retries stack with the SDK’s own

Two retry layers

The OpenAI and Anthropic SDKs already retry twice by default (max_retries=2). With parry on top, a single request can be attempted max_attempts × (max_retries + 1) times, which is 9 with the defaults.

Pick one layer to own retries:

# parry is the only retry layer (recommended: consistent policy, counted in stats)
client = openai.OpenAI(max_retries=0)

# or keep the SDK's retries and switch parry's off
parry.install(retry=False)

Retries and the other features

• Each retry takes a fresh rate-limit slot.
• Every retry is counted in stats.retries, and calls that finally fail in stats.errors.
• The budget is checked once, before the first attempt.

Standalone use

from parry import Retry

retry = Retry(max_attempts=5, base_delay=0.25)
result = retry.call(flaky_function, arg, key="value")   # sleeps with the GIL released
retry.delay(3)                                           # inspect the schedule
Retry.is_retryable("RateLimitError", 429)                # True