Busy ran the business and could not be switched off, so nothing was ever migrated — a bridge let both systems run at once.
Busy ERP was the accounting system the company actually ran on. Stock, balances and master data lived there, and every plan for replacing it began with a cutover weekend.
A cutover has one failure mode and it is total: if Saturday goes wrong, nobody can invoice on Monday, and the rollback is a restore from backup plus a lost weekend of transactions. For a business that dispatches daily, that risk was never worth taking for any amount of architectural tidiness.
Every version of the migration plan had the same shape: freeze, move, verify, switch. I killed all of them. The plan is only sound if verification is perfect, and verification is never perfect on data that has been accreting inside an accounting system for years. Refusing to migrate was the decision the whole project turned on.
One service consumes Busy's events and speaks to the modern stack. Nothing else in the system knows Busy exists. That boundary means the legacy schema cannot leak into new code, and either side can be redeployed or rolled back without the other noticing.
The sync scripts take --dry-run, --force and environment flags, and the default is the safe one. Every run against production was printed and read before it was executed. Nothing shipped that only worked if it worked.
Gzip-compressed events, connections that drop without notice, and no replay. Auto-reconnect and idempotent handling were not optional extras, they were the baseline for it working at all.
Both systems can change the same master record. Deciding which side wins, per entity type, was more of the design work than the transport ever was.
The honest outcome is that the bridge is permanent infrastructure, not a temporary migration aid. Planning it as temporary would have produced something too fragile to still be running.
The best migration is often no migration. Two systems side by side beats one risky switch.
If a plan's rollback is restore-from-backup, it is not a plan, it is a hope.
Fence legacy systems behind one service. The moment two places know about the old schema, you own it forever.