Why it matters
Data loss from bad Cassandra backup practices is a common enterprise incident. Getting backups right is basic operational hygiene.
The architecture
Snapshot: nodetool snapshot creates hard links to current SSTables in snapshots/ directory. Instant, uses no extra space initially.
Incremental backup: on flush, new SSTable hard-linked to backups/ directory. Enables PITR-style recovery.
How it works end to end
Off-cluster storage: after snapshot, copy files to S3 or another DC. rsync or dedicated tools like Cassandra Backup Solutions.
Restore per-node: copy files to data directory, then use nodetool refresh or restart.
Cluster-wide consistent restore is hard: snapshots aren't taken at the same instant across nodes. Coordinate via out-of-band tools or accept eventual consistency after restore.