August Recap: AI Is Reshaping the DBA Role

A Month on AI and the DBA Role

August was spent on one question from many angles: what does AI actually change about database migration work, once you strip out the claims nobody can substantiate? A few threads ran through everything.

Rules Where the Answer Is Known

The recurring technical conclusion. Type mapping, function swaps, DDL shape, construct detection — these are finite, known problems, and a deterministic rule that is always right beats a model that is usually right. A model that “usually” spots an autonomous transaction is strictly worse than a pattern that always does.

Models earn their place at interpretation: explaining an undocumented package, summarising a query plan, proposing a redesign where no equivalent exists.

The Dangerous Failures Are Quiet

Every failure case worth writing about shared a shape: the migration appeared to succeed. Row counts matched. Nothing errored. An empty string became a real value, a timestamp lost its time component, a sequence sat at its initial value until the first insert collided.

We included one of our own — a transpiler emitting routines with unbalanced dollar-quoting, invisible to a large test suite because every test checked fragments rather than whether the output was structurally whole. The lesson generalises past our bug: assert the property you actually care about.

Verification Is Becoming the Product

Generating a migration is commoditising; proving it correct is not. That is why row counts are table stakes and insufficient, why column-level checksums matter, and why per-partition proof exists — to localise a mismatch in a large table rather than merely announce one.

Local Is a Governance Answer

Schema and query text describe a business in detail. Running inference in-process makes “where does this go?” a short conversation, and small fine-tuned models are now good enough for this narrow domain to make that a real option rather than a compromise.

What We Declined to Claim

We published the conversion rate — 95.5% on a reproducible 4,000-object benchmark — because it is measured. We did not publish an end-to-end time saving, a cost saving, or an ROI multiple, because those depend on inputs only the reader has. Instead each of those posts gave the model and asked you to fill it in.

That is a worse marketing decision and a better engineering one, and it is the thread we would most like to keep.

The Role Itself

The DBA role shifts the way it shifted when backups were automated: away from executing the procedure, toward specifying it, verifying it, and owning the decision. The valuable skill stops being “I can write the conversion” and becomes “I can tell whether this conversion is right” — which needs more knowledge of both engines, not less.

September moves from the AI layer to the operational one: running PostgreSQL well after the migration lands.