Zero lost characters

Each column declared its own character set — latin1, latin2, UTF-16 — and a single users table mixed two of them: latin2 for first and last names, latin1 for everything else. Read naively through a UTF-8 connection, a name like Nicolás comes out as Nicolás — or, worse, characters vanish without a trace.

People were spread across three sources — personal records, logins and teaching assignments — with no shared key, and student numbers were unique only within a degree programme.

A sync engine in Node.js that reads the legacy database, plans every write as a pure transformation, checks the whole plan against a quality gate and only then writes — in foreign-key order, one transaction per logical unit, so a crashed run can simply be run again.

It is insert-only: it never updates or deletes. A mapping table (entity, legacy key → new id) remembers what has already moved. For each item: if its key is in the map, skip it; if a row with the same natural key already exists — an email, a user and course, a catalog code — adopt it into the map; otherwise insert it. That middle rule let the engine take over a production database that earlier scripts had already filled.

Three modes for three moments: a preview that writes nothing, a bootstrap that only builds the map, and real runs that require an explicit --allow-write. A wrapper for the operator opens the connections, runs it and prints the outcome in plain language.