Hourly billing pays for the slow fix. We price the quiet one.

We start where the noise is loudest: pg_stat_statements sorted by total_exec_time, the MySQL slow log, wait events piling up at the same minute every evening, a buffer cache hit ratio that has drifted three points since March.

Then we read plans line by line, the way you’d run a thumb down a fretboard feeling for the high spot — a nested loop over ninety thousand rows, a sort spilling to disk because work_mem is still the default, an index that hasn’t been scanned once since it was created.

Three weeks with our hands on the strings. We restore a production-sized clone, time every target query cold and warm, and rewrite the twelve that cost you most — a correlated subquery turned into a lateral join, an OR split into a UNION ALL the planner can actually use, an N+1 folded back into one round trip.

Every change arrives as a pull request with its timed plan pasted in the description, so your reviewer hears the difference before they approve it.

Migrations go wrong in the quiet parts — a collation that sorts “é” one place on the old box and another on the new one, a sequence that didn’t follow its table, a large object nobody remembered was there. We inventory every one before replication starts.

Then we rehearse. Twice, on a clone, with a stopwatch. By cutover night the switch is a connection-string flip your engineers have already done once with their own hands.

Databases drift the way guitars drift between seasons. Tables triple, a new feature ships its own query, someone adds a column with a default and a lock nobody noticed. Once a quarter we re-run the original audit queries against the baseline we left you and listen for anything that’s started to hum again.