Replication-manager remembers what each logical backup looked like when it was written or registered. If the backup's files change afterwards, by hand or by another tool, replication-manager stops using that backup until you register the change. Your files are never deleted or modified: the backup stays recoverable, but replication-manager will not reseed from it on its own.
This page explains what counts as a change, what is refused, and how to edit a backup on purpose, for example removing tables from a splitdump or mydumper folder.
When a logical backup completes (mysqldump, splitdump, mydumper), or when you register a backup found in the backup folder, replication-manager records a fingerprint of its files:
The change time is there because the modification time can be put back (touch, cp -p, rsync -t), the change time cannot: the kernel updates it on every change to the file, so a backup replaced by a copy with the same size and date is still detected.
Physical backups (mariabackup, xtrabackup) have no fingerprint and are not checked. Off Linux no fingerprint is taken.
Anything that makes the files differ from the fingerprint:
chown, chmod), an ACL, or an attribute restore: these update the change time without touching the content.Development builds before #2114 changed the owner of every file in replication-manager's working directory at each start, so every registered backup was reported changed after each restart. Replication-manager now leaves the files already owned by its user alone.
The Maintenance tab marks the backup Changed since registered: register the change first, and the cluster raises WARN0249 per server, listing each changed backup and how it changed (sizes, file counts, modification time, or "replaced or rewritten, same size and modification time").
When no other logical backup is usable, WARN0190 (no logical backup available for rejoin) follows.
Until the change is registered, replication-manager refuses that backup for:
It does not stop you from restoring it yourself: the files are untouched, myloader or mysql run by hand work as before.
Moving files out of a dump is a normal operation, for example:
The steps:
POST /api/clusters/{cluster}/backups/{id}/register.If Register refuses it, the reason is shown on the backup. Unregister removes the backup from the catalogue without touching its files.
To go back to the original instead, restore it from its archive and register it: the restored file is a new file, so it is a change too.
The content did not change, but the backup is reported changed all the same. Register it: replication-manager reads it again and records it.