I have experience in migrating large DBs with replication and the article not discussing write blocks made my ears perk up as well.
replyAside from the blocking you mentioned during the initial snapshot, you'd need to block writes to the old DB before the cutover as well. There's no way to guarantee in-flight writes to the old DB aren't lost when promoting the replica to a primary otherwise. I'm surprised the author didn't go into more detail here. Maybe it was fine given their workload, but the key issue I see is that they promoted the new DB to a primary before stopping the old application. During that gap, any data written to the old DB would be lost.
Correct me if I'm wrong, but done with a proxy in-between that can "pause" requests, you could have done the move with 0 seconds and no rejected requests, and I don't think mydumper/myloader/xtrabackup matters for that. The "migration" would be spinning up a new database, making it catch up, then switching over. If you can pause/hang in-flight requests while switching, not a single one needs to fail :)
replyThe "making it catch up" is the tricky part. You need an initial backup for that. xtrabackup can take that backup "hot" without blocking read/writes. mysqldumper will block writes for whatever time that initial backup takes, for 2TB of data that's going to be hours.
replyOnce you have that initial back up you can set your replica and make it catch up , then you switch. I choose to take the few seconds of downtime doing the switch because for my use case that was acceptable.
Isn't that just a flag? "--lock-tables=false", alternatively --single-transaction for InnoDB.
replyIf you want a consistent backup that you can use to setup a replica you need to block writes while the backup is taken, take the backup while the database is shutdown OR use xtrabackup.
reply