I Made Kimi K3 My Database Analyst for a Week: 47 Tables, Real Queries, a $2.17 Bill

Why Database Work Is the Real Test

Everyone benchmarks models on homework problems. I run a small data consultancy, and my actual week looks like this: a client's warehouse with 47 tables and comments from three different contractors, a retention question from marketing, and a query that suddenly takes 40 seconds because someone doubled the row count. SQL is the perfect AI test because it has ground truth — the query either returns the right number or it doesn't.

So for one week I did all of it with Kimi K3: the schema archaeology, the query writing, the optimization. No copilot autocomplete, no previous session memory — fresh context each morning. Full honesty on the good and the embarrassing.

Feeding It 47 Tables

Day one, I ran pg_dump --schema-only and pasted the whole thing — about 55,000 tokens of CREATE TABLE statements, constraints, indexes, and the least helpful comments I've ever written. K3 swallowed it in one prompt without truncation warnings.

The first thing I asked for was a relationship map in plain English. It produced an accurate walk from users to orders to line_items to products, and — the part that impressed me — it flagged that orders.customer_id and orders.user_id looked redundant. They are. A contractor added user_id two years ago "for convenience," and both are populated in production with a 0.3% drift. K3 spotted the type of mess I charge clients to find. That single catch paid for the test many times over.

The Queries That Earned Their Keep

Three jobs stood out over the week:

• Cohort retention. I asked for weekly retention by signup cohort with a twist — only counting users who performed a "key action" event. K3 wrote it with a GENERATE_SERIES-based week spine, which is how my textbooks do it. I actually ran it and the numbers matched my old dashboard exactly.
• JSONB unnesting. Our events table hides custom properties in a JSONB column. I described the shape in one paragraph (no DDL — the keys are schema-less) and asked to count event types per user per day. First try, it nailed the jsonb_array_elements + LATERAL pattern, including the edge case of events with missing keys — COALESCE handled it.
• The 40-second query rewrite. A monthly revenue rollup joined four tables with OR conditions in the WHERE clause. K3 rewrote it to pre-aggregate before joining, replace the ORs with a UNION branch, and add a partial index suggestion on status = 'completed'. Runtime: 40s → 1.3s. My client thinks I'm a genius; K3 did it in one message while I ate a sandwich.

Where It Stumbled

Two failure classes, both worth internalizing before you trust it with production changes. First, dialect drift: despite me stating Postgres in the prompt, it produced MySQL DATE_FORMAT masks in two answers when I asked follow-up questions a day later. Fresh context each morning is my own workflow, but it means the dialect reminder has to be part of every prompt — it doesn't carry over, and it doesn't ask.

Second, confident misreading of business logic. We had a refunds table that's a soft ledger with reversals — a refund row and a reversal row can cancel out. I asked for "total refunds by month" and got a number that double-counted net-zero pairs. K3 can't know the semantic rule unless I state it. When I added one sentence — "reversal rows cancel matching refund rows" — it rewrote the query correctly, including the self-join pattern I wanted. The lesson: the model owns SQL syntax; you own the meaning of your data.

The Workflow I Kept

After the week, I kept three habits. Every analytical session opens with the schema dump as the FIRST message — it becomes a cached prefix and cuts costs meaningfully. Every business rule I care about goes into a standing "data dictionary" paragraph I paste after the schema. And nothing gets run in production until I've read it out loud; the 85% first-try accuracy sounds great until you remember what the remaining 15% does to a revenue number someone presents to a board.

Cost total for the week: $2.17. Compare that to my own hourly rate and you can guess whose invoice I'm not competing with. If you're weighing K3 for engineering work generally, my data science workflow piece covers pandas and notebooks; and the $5 coding test covers the rest of the developer surface.