Robbie Ostrow, OpenAI
Developers interact with codebases one commit at a time, but the distributed systems we manage change much more chaotically. Services are constantly deploying at different versions and rollouts are never instant. A simple change to a wire format can break inflight requests, data in queues, caches, or databases.
Data formats like Protocol Buffers exist to solve some of these problems, but they aren’t expressive enough to encode higher-level invariants. Keeping track of breaking changes and managing multi-step deployments by hand is too much overhead for any engineer.
In this talk, we get a bit more formal about what it means to actually make a “breaking change” and introduce simple tooling that shifts rollout safety from individual developers into the system, while enabling stricter API contracts that prevent entire classes of failure.
Robbie Ostrow works on infrastructure and reliability at OpenAI, where he spends far too much time trying to formally express the invariants he expects from his systems. Before OpenAI, Robbie led engineering teams at Q Bio and Vanta, where he preached the same ideas at much smaller scales. When he isn’t thinking about type systems or breaking changes, you can find him in San Francisco playing word games or hunting down merch from failed companies.
author = {Robbie Ostrow},
title = {Escaping Version Skew: Formalizing Compatibility in a World of Partial Rollouts},
year = {2026},
address = {Seattle, WA},
publisher = {USENIX Association},
month = mar
}
