Backups and recovery

A backup only counts if it runs, is protected, and can actually be restored in time.

// TECH FOR GRC

Two numbers that define recovery

  • RPO (recovery point objective): how much data the business can afford to lose. Daily backups mean an RPO of about 24 hours.
  • RTO (recovery time objective): how quickly systems must be working again.

Both are usually promised to customers, so auditors test against them.

Backup is not replication

Replication copies data to another place in real time, which helps if a data centre fails. But if data is deleted or encrypted by ransomware, replication copies the damage too. Backups are point-in-time copies you can go back to.

Protecting the backups

Attackers go after backups first. Immutable backups, such as AWS Backup Vault Lock or S3 Object Lock, cannot be deleted or changed, even by an administrator, for a set period.

What this means for evidence

  • The backup plan: what is backed up, how often, and kept for how long.
  • Job history for the period: completed, and failures handled.
  • A restore test with the date, the system, the time taken against the RTO, and a sign-off. This is the evidence most often missing.
// WHAT IT LOOKS LIKE AS EVIDENCE
Backup jobs running and monitoredAWS Backup, Azure Backup or GCP Backup and DR · proves A1.2, A.8.13
AWS BackupSAMPLE
AWS Backup › Jobs › Backup jobs

Backup jobs: last 7 days

complyoo-prod (4821 0093 1176)Exported 2026-10-011
ResourcePlanStartedStatus2Vault
rds:complyoo-maindaily-35d2026-09-30 01:00Completedprod-vault (locked)
efs:uploadsdaily-35d2026-09-30 01:00Completedprod-vault (locked)
rds:complyoo-maindaily-35d2026-09-29 01:00Failed, retried 03:00 Completedprod-vault (locked)3
  1. 1Export date and account.
  2. 2Status per job.
  3. 3A failure that was noticed and retried. Evidence of monitoring, not a problem.

Weak version: The backup plan configuration with no job history. It shows backups are planned, not that they ran.

// FROM A REAL ACCOUNTAWS console, AWS Backup · identifiers blurred
A backup plan: every day at 20:00 UTC, kept for 35 days, in a named vault. The schedule preview lists the next ten runs, which answers how often before anyone asks.
StrongA backup plan: every day at 20:00 UTC, kept for 35 days, in a named vault. The schedule preview lists the next ten runs, which answers how often before anyone asks.
The job that actually ran: Completed, with the resource and the time. A plan proves intent; the job list proves it happened.
StrongThe job that actually ran: Completed, with the resource and the time. A plan proves intent; the job list proves it happened.
A restore test with result and sign-offTicket or DR report · proves A1.3, A.8.13, A.5.30
Restore test recordSAMPLE
Jira › OPS-318

OPS-318 Quarterly restore test: complyoo-main database

Ticket closed 2026-09-25
Performed
2026-09-24 by maria.s1
Restored from
Snapshot of 2026-09-23 01:00 into an isolated account
Result
Restored in 41 minutes. Row counts matched. RTO target 4 hours met.2
Signed off
David E. on 2026-09-253
  1. 1Who did it and when.
  2. 2What was restored, how long it took, and whether data checked out, against the RTO.
  3. 3A reviewer signed off.

Weak version: "Restore tested, all good." No system, time taken or validation.

// FROM A REAL ACCOUNTAWS console, AWS Backup and DynamoDB · identifiers blurred
A restore job from that backup into a new table: Completed in 6 minutes. That number is your measured recovery time, not the one in the policy.
StrongA restore job from that backup into a new table: Completed in 6 minutes. That number is your measured recovery time, not the one in the policy.
Then check the data came back: the restored table holds order-1001 with its amount. A restore that finishes but returns nothing is not a passed test.
StrongThen check the data came back: the restored table holds order-1001 with its amount. A restore that finishes but returns nothing is not a passed test.
// REMEMBER THIS

An untested backup is a hope, not a control.

// CHECK YOURSELF

1. The business can lose at most 4 hours of data. That is the...

2. Why doesn't replication replace backups?

3. Which evidence is most often missing?