Fixing money calculations with no automated tests

What arrived

• A purchased exchange script with a couple of weeks of one developer's changes on top.
• No automated tests and no migrations. The schema came from sequelize.sync the first time the app started against an empty database.
• Nothing stated what any money path was supposed to do. The running code was the only definition of correct.

How changes were checked

Each change was tried by hand in a staging environment before release. The regressions below were found there.

The order of fixes

• Day 3: leverage, which the admin panel saved twice and nothing read.
• Day 14: leverage applied to the margin charged when a trade opened.
• Day 20: one margin and profit formula across the endpoints; liquidation reimplemented as a 30-second job.
• Day 25: the single frontend store split into five, across 96 files.
• Day 29: order execution, position close and every timed job wrapped in a database transaction. None had been.

Features in the same period

Crypto deposits in the first week, fiat deposits after, then TradingView charts with a REST price feed behind a three-second cache.

Regressions

• The day-20 rewrite read an order's metadata as an object where it was stored as a JSON string, and limit orders got zero margin reserved. It was found and fixed the next day.
• A bonus-balance feature on day 26 left cancelled orders holding their margin. Release on cancellation came three days later, in the change that wrapped it in a transaction.

Left open

Two of the three timed jobs, margin calls and stop-loss, were registered under the same name at handoff.