Unit test a transcription retry loop with a fake 429 in Python

You can test a Sume transcription retry loop without the network by passing the HTTP call and the sleep function in as arguments. A fake post returns (429, {"retry-after": "2"}) and then (202, {}); the test asserts that the loop slept 2.0 seconds and sent the same Idempotency-Key both times. Those two assertions are the contract in Errors and rate limits and Jobs and results: honor retry-after, and retry with the same key so no second paid job is created. Read 2026-10-06.

What to assert

A retry loop has three behaviors worth pinning down. Anything beyond that is testing the standard library.

The code and its tests

Save it as ut2.py and run python -m unittest ut2. Both tests pass in a few milliseconds because nothing sleeps.

import time, unittest


def submit_with_retry(post, key, sleep=time.sleep, tries=5):
    for _ in range(tries):
        status, headers = post(key)
        if status == 429:
            sleep(float(headers.get("retry-after", "5")))
            continue
        return status


class RetryTest(unittest.TestCase):
    def test_retry_keeps_the_key(self):
        keys, naps = [], []
        replies = iter([(429, {"retry-after": "2"}), (202, {})])

        def post(key):
            keys.append(key)
            return next(replies)

        self.assertEqual(submit_with_retry(post, "stt-1", sleep=naps.append), 202)
        self.assertEqual((keys, naps), (["stt-1", "stt-1"], [2.0]))

    def test_gives_up(self):
        post = lambda key: (429, {})
        self.assertIsNone(submit_with_retry(post, "k", sleep=lambda s: None, tries=3))

Wiring the real call in

In production, pass a post that wraps requests.post and returns (response.status_code, response.headers). The loop does not care where the numbers come from, which is what makes it testable.