Backups

Backups Exist for Three Different Disasters

 1. OOPS:        someone dropped the table / bad migration
 2. CORRUPTION:  bad data replicated everywhere before noticed
 3. LOSS:        region gone, ransomware, storage failure

 each demands different backup properties:

 OOPS:        point-in-time granularity (restore to 14:32, 
              NOT last night's snapshot)
 CORRUPTION:  RETENTION going back far enough to pre-date it,
              + awareness corruption exists (detection!)
 LOSS:        geographic separation (different blast radius)

The Strategy Stack

 FULL backups:      everything, periodic (weekly typically)
 INCREMENTAL:       changes since last backup (daily/hourly)
 CONTINUOUS/WAL:    transaction log shipping → PITR
                    (point-in-time recovery to any second)
 SNAPSHOT:          volume-level, fast, coarse

 production pattern:
   weekly full + hourly incrementals + continuous WAL
   = restore to any minute within retention window

 retention tiers:
   24h:   hourly      (oops recovery)
   30d:   daily       (corruption discovery window)  
   1y+:   monthly     (compliance/legal)

The Rules That Separate Real From Theater

 □ 3-2-1: THREE copies, TWO media/systems, ONE offsite.
   (the offsite copy survives ransomware/ransomware-class events)
 □ ENCRYPTED at rest AND access-controlled — backups are
   a data-leak surface; least privilege applies
 □ IMMUTABLE copies where possible (object-lock):
   attackers and fat-fingered admins can't delete them
 □ TESTED restores on schedule (own lesson — untested
   backups are hopes, not backups)
 □ AUTOMATED verification: checksums, restore-drill reports
 □ APPLICATION-CONSISTENT: database-aware snapshots, not
   blind disk copies mid-write (torn pages = corrupt restore)

Sizing the RPO Conversation

Backup approach	Typical data loss window
Nightly full	up to 24h of writes
Hourly incrementals	≤1h
Continuous WAL/PITR	seconds
Cross-region replication	seconds–minutes (DR lesson)

 RPO requirement drives cost nonlinearly:
 "lose nothing ever" = synchronous replication pricing.
 most systems: PITR for oops + cross-region async for disaster.

Interview Framing

“What happens when someone drops the users table?” scored answer: PITR restore-to-timestamp flow (not nightly-snapshot despair), retention math vs corruption-discovery delay, 3-2-1 with immutable tier, consistency guarantees of the backup mechanism named. The follow-up hiding here is always “when did you last TEST it” — have that answer ready before they ask.