Engineering

Distributed systems and persistence

AI features still depend on ordinary backend engineering. State has to persist, retries have to be safe, and partial failures need a way to recover.

Keeping state clear

I want a data model that makes it clear what is stored as fact, what can be recomputed, and where temporary inconsistency is acceptable.

Handling partial failure

A dependency can time out while the rest of a workflow succeeds. Duplicate messages and repeated calls need deliberate handling so a retry doesn't create a second result.

Open questions

  • What is stored as fact, and what is recomputed?
  • What happens if this step runs twice?
  • Which failure is visible to the person who has to repair it?