Canary 10% of video jobs to Sume before cutover: sticky bucketing

Route a fixed percentage of video jobs to Sume by hashing a stable key, such as the customer id, into a 0-99 bucket and sending buckets below the threshold to the new backend. The same customer always lands on the same side, so you compare like with like and a retry never hops backends mid-job. This matters most for teams whose previous vendor is gone: OpenAI lists the Sora API shutdown date as 2026-09-24 (read 2026-10-03), so the canary compares Sume against a fallback rather than against the old vendor.

Sticky bucketing

Hash the key, not a random number per request. A random draw would send one customer's retry to a different backend and defeat the idempotency key, which only dedupes within one API.

import asyncio
import hashlib

def bucket(key: str) -> int:
    return int(hashlib.sha256(key.encode()).hexdigest()[:8], 16) % 100

def backend(customer_id: str, sume_percent: int) -> str:
    return "sume" if bucket(customer_id) < sume_percent else "fallback"

async def main() -> None:
    for pct in (10, 50, 100):
        share = sum(backend(f"cust-{i}", pct) == "sume" for i in range(1000))
        print(pct, share)

asyncio.run(main())

What to compare during the ramp

Compare terminal outcome rate, time to terminal and cost per accepted clip, and treat a quality review by a person as a metric too. The Sume side already reports status and error codes in a fixed vocabulary, so the comparison needs only that your fallback adapter maps its own errors into the same buckets.

Keep the fallback

Reaching 100% should not delete the interface from the previous section. A vendor that shut down once is the best argument for keeping a second adapter warm.