Why do scale-out databases still require backup despite built-in replication and redundancy?

Scale-out databases are designed to keep systems running. They replicate data across nodes, racks, or even regions. That protects against hardware failure and improves uptime. But it does not protect against bad data.

If data is corrupted, deleted, or modified incorrectly, that change is replicated across the system just as efficiently as valid data. In other words, replication spreads problems as well as it spreads availability.

There is also a human and application error factor. Accidental deletions, faulty updates, or misconfigurations are common. Replication cannot distinguish between intentional and unintentional changes. Once the mistake happens, it is propagated everywhere.

Another gap is lack of point-in-time recovery. Replication keeps systems synchronized in the present. It does not give you the ability to go back to a known good state from yesterday or last week. Backups and snapshots provide that historical recovery capability.

Security is another concern. In scenarios like ransomware or malicious activity, replication will often replicate encrypted or compromised data across all nodes. Without isolated backups, recovery options become extremely limited.

There is also a compliance and audit requirement. Many industries require long-term data retention and the ability to recover historical states. Replication does not address this because it focuses on current state consistency, not historical preservation.

Looking at the evolution of infrastructure helps clarify this. RAID, clustering, and replication all improved availability over time. None of them removed the need for backup. Scale-out databases follow the same pattern.

Snapshots and backups add capabilities that replication does not provide:

point-in-time recovery, isolation from source failures, and protection against logical errors.

The practical takeaway is simple.

Replication keeps systems running.

Backup ensures you can recover when something goes wrong.

You need both, because they solve different problems.