Ghost table migration lock contention

A 'zero-downtime' column-type migration (gh-ost) on a 400M-row MySQL 8 users table was kicked off at peak. Twelve minutes in, the API's p99 went from 80ms to 9s and the error rate hit 30%. Dashboards: MySQL Threads_running pinned at the connection-pool ceiling, History list length (undo log) climbing fast, replica lag on the read replicas rising past 200s, the gh-ost ghost-table copy at 60%. The on-call before you already tried killing the app pods, which didn't help. What's happening, how do you stabilize, and how do you finish or unwind the migration without losing rows?