backups（backups 效率）

The Only Rule That Matters

• Untested backups are not backups — schedule regular restore tests, not just backup jobs
• Test restores to different hardware/location — validates both backup and restore procedure
• Time the restore — know how long recovery actually takes before disaster strikes

3-2-1 Rule Violations

• Same disk as source data = not a backup — disk failure loses both
• Same server as source = not a backup — ransomware/fire/theft takes both
• Same cloud account = risky — account compromise or provider issue loses both
• Cloud sync (Dropbox, Drive) is not backup — syncs deletions and corruption too

Ransomware Protection

• Backups accessible from production get encrypted too — air gap or immutable storage required
• Append-only/immutable storage prevents deletion — S3 Object Lock, Backblaze B2 with retention
• Offline rotation (USB drives, tapes) for critical data — can't encrypt what's not connected
• Test restoring from immutable backup — verify ransomware can't corrupt the restore process

Database Backup Traps

• File copy of running database = corrupted backup — use pg_dump, mysqldump, mongodump
• Point-in-time recovery needs WAL/binlog archiving — dump alone loses recent transactions
• Large databases: pg_dump locks tables — use pg_basebackup or logical replication for zero downtime
• Test restore to different server — verifies backup is self-contained

Incremental Backup Pitfalls

• Incrementals depend on chain — one corrupted backup breaks all following
• Long chains slow restores — schedule periodic full backups
• Deduplication saves space but adds complexity — single repo corruption affects all backups
• Verify backup integrity regularly — bit rot happens, checksums catch it

Retention Mistakes

• No retention policy = storage fills up — define and automate cleanup
• Too aggressive retention = can't recover old corruption — keep monthlies for a year minimum
• Legal/compliance requirements may mandate retention — check before setting policy
• Grandfather-father-son pattern: daily/weekly/monthly tiers

Filesystem Traps

• Permissions and ownership often lost — verify restore preserves them, or document expected state
• Symlinks may not backup correctly — some tools follow, some copy link, test behavior
• Sparse files may inflate — 1GB sparse file becomes 1GB actual in backup
• Extended attributes and ACLs — not all tools preserve them

Cloud and Remote

• Encrypt before upload — cloud provider breach shouldn't expose your data
• Bandwidth costs add up — initial seed via physical drive for large datasets
• Region matters for disaster recovery — same region as production doesn't survive regional outage
• Egress fees can be brutal — know restore costs before emergency

Tool-Specific

• rsync --delete on wrong direction destroys source — always double-check source/destination
• restic/borg need repository password — lose it = lose all backups, no recovery
• Tarball without compression: faster, but larger — choose based on CPU vs storage tradeoff
• Snapshots (LVM, ZFS, cloud) are not backups — same storage system, same failure domain

Documentation

• Document restore procedure — you won't remember under pressure
• Store procedure outside the backup — printed, different system, password manager
• Include credentials, paths, expected time — everything needed to restore at 3am