Backup vs Replication

Choose backup, snapshots, logs, replication, and immutable copies based on outage, corruption, deletion, ransomware, retention, RTO, and RPO requirements.

Replication keeps another copy of current state available. Backup preserves recoverable state from an earlier point. Both copy data, but the time relationship and failure isolation are different.

That distinction matters because many incidents damage current state rather than destroy one storage device. Accidental deletion, application bugs, malicious encryption, credential compromise, and silent corruption may be copied perfectly to every live replica.

Compare the mechanisms

Mechanism Main purpose Strength Limitation
Synchronous replication Keep replicas at the same committed state within the protected scope tight RPO; rapid local failover latency and quorum trade-offs; logical errors propagate
Asynchronous replication Maintain a geographically or operationally separate recent copy lower write latency; broader distance replication lag and possible data loss on promotion
Snapshot Capture storage state at a point in time fast creation and useful rollback point application consistency and storage dependency must be verified
Full backup Preserve a complete recoverable data set simple recovery base longer copy and restore time
Incremental or differential backup Capture changes relative to another backup lower transfer and storage cost restore chain and catalog integrity become dependencies
Transaction or write-ahead log archive Preserve changes between base recovery points finer point-in-time recovery requires a valid base copy and correct replay
Versioning or history Retain prior object or record states useful for deletion and overwrite retention, lifecycle, and administrative protection still matter

Replication supports availability. Backup supports historical recovery. A resilient data design usually needs both.

Match the control to the failure

Failure Replication helps? Backup or history helps? Important caveat
Disk or node loss Yes Yes Replica failover is usually faster
Zone or site loss Yes, if outside the failed scope Yes, if accessible elsewhere Network and control dependencies must also survive
Accidental deletion Often propagates the deletion Yes Recovery point must predate the deletion
Application corruption Often propagates bad state Yes Detection delay determines required retention
Ransomware May replicate encrypted data Yes, if copies resist the compromised credentials Restore into a clean environment
Long-term retention Not usually the right control Yes Legal holds and deletion policy may apply
Very low interruption target Yes Usually too slow alone Promotion and traffic movement still affect RTO

Independence determines backup value

A copy is more useful when it is independent across several dimensions:

  • Failure domain: it does not share the same host, rack, zone, site, or storage control plane.
  • Administrative boundary: the same compromised identity cannot alter production and every recovery copy.
  • Time: historical points survive long enough to predate delayed corruption.
  • Technology: one software defect or format failure does not invalidate every copy.
  • Access path: recovery operators can reach the copy when primary identity, DNS, or network services are impaired.
  • Encryption and keys: required keys are protected and recoverable without exposing them broadly.

“Stored in another region” addresses only one dimension. A cross-region copy controlled by the same compromised account may not provide cyber recovery.

Immutability and isolation

Immutable or write-once retention prevents protected copies from being changed or deleted during a defined period. It is valuable against destructive mistakes and malicious actions, but it is not magic:

  • retention must exceed likely detection time
  • privileged bypass paths must be understood
  • encryption keys and metadata must remain available
  • stored data must be scanned or validated before recovery
  • restore capacity and clean credentials must exist
  • legal and privacy deletion obligations still apply

Air-gapped is often used loosely. Be precise about whether the copy is physically offline, logically isolated, administratively separate, or merely in another account or project.

Crash-consistent vs application-consistent copies

A storage snapshot may capture blocks as if power were removed suddenly. A database or distributed application may require coordinated flushing, transaction logs, quiescing, or service-native backup APIs to produce an application-consistent recovery point.

Before relying on a snapshot, ask:

  • Which writes were acknowledged?
  • Are related volumes captured consistently?
  • Can the application replay its logs after restore?
  • Are external objects, queues, and configuration included?
  • Does the restored system preserve referential integrity?

A successful snapshot operation proves that blocks were copied, not that the business service can recover correctly.

The restore chain is a dependency graph

Point-in-time recovery might require:

  1. Locate the correct full or base backup.
  2. Verify integrity and decrypt it.
  3. Restore the base data.
  4. Apply differential or incremental backups in order.
  5. Replay transaction logs to the selected recovery point.
  6. Reconnect identity, application, queue, and storage dependencies.
  7. Validate consistency before accepting traffic.

Missing one log segment, catalog entry, key, artifact, or credential can break the chain. Recovery tests should validate the complete sequence, not merely list stored objects.

Backup design checklist

Decision Question
Scope Which data, configuration, identities, keys, artifacts, and infrastructure definitions are required?
Frequency Can the produced recovery points meet each flow’s RPO?
Retention How long might corruption remain undetected, and what obligations apply?
Isolation Which failures and credentials can affect both production and copies?
Integrity How are incomplete, empty, corrupted, or malicious copies detected?
Restore Where will recovery run, and is sufficient capacity available?
Validation What evidence proves application and business correctness?
Disposal How are expired copies deleted safely and verifiably?

Common reasoning errors

Error Better rule
“Three replicas means three backups.” Replicas usually preserve current service state; backups preserve independent historical states.
“The backup job is green, so recovery is ready.” Validate restoration, keys, dependencies, timing, and data correctness.
“Cross-region means ransomware-safe.” Administrative and credential isolation matter as much as geography.
“Snapshots are always application-consistent.” Verify the application’s consistency and log-replay requirements.
“More frequent backups guarantee a tighter RPO.” Completion, transfer, integrity, catalog, and restore capability determine the effective RPO.

Continue with Failure Domains and Blast Radius to evaluate whether copies and live replicas actually fail independently.

Knowledge check

Loading quiz…

Further reading: Google SRE data integrity and NIST contingency planning guidance.

Revised on Friday, September 11, 2026