Ops Engineering Shift

AI writes a working pipeline in 10 minutes. Nobody has asked it to be on call for one.

I’ve watched this play out in a few teams this year. Someone prompts a model, gets a clean dbt model plus an ingestion job, and it runs green on day one.

Then week three happens. A source sends yesterday’s file twice, and March needs a backfill that doesn’t double-count it. A late partition lands after the dashboard already refreshed.

None of that shows up in the generated code, because none of it was in the prompt. It lives in the person who’s seen it go wrong before.

So the job is drifting toward operations: making loads idempotent, and knowing what’s downstream before you touch anything. That’s what I’d test in an interview now. Hand the candidate a pipeline that ran twice and ask how they’d clean it up.

Writing SQL still matters. It’s also the part a model does reasonably well, and I honestly don’t know yet how far that goes.

When did your team last rerun a load, and did it come out right the first time?