Disaster recovery is the planned restoration of systems and data after a major outage, and a backup is where that restoration starts. In the cloud, the provider keeps the platform running, but your team decides what gets backed up, how quickly services must return and how recovery is tested. This guide covers the concepts, strategies and rules.
Quick Overview
What you will learn:
- What RTO, RPO and maximum tolerable downtime mean
- Which recovery strategies public clouds support, and what shorter recovery targets cost
- How to protect backups from ransomware and human error
- What NIS2 and DORA require
Who this article is for:
- Cloud architects, administrators and SRE teams
- People responsible for business continuity and IT security
Reading time: 8 minutes
Where the Provider’s Responsibility Ends
Microsoft’s article on shared responsibility for reliability (learn.microsoft.com) draws the line clearly. Microsoft is solely responsible for core platform reliability. It also provides reliability features such as availability zones and backups, but you are responsible for choosing and configuring them: checking that backups are actually enabled, picking the right service tier and designing applications that tolerate faults. A managed service does not, by default, mean a recoverable workload.
Key Terms: RTO, RPO and the Recovery Plan
NIST SP 800-34 Rev. 1, the “Contingency Planning Guide for Federal Information Systems” (csrc.nist.gov, published May 2010 and updated in November 2010), gives the shared vocabulary. Three values come out of a business impact analysis (BIA):
- MTD (maximum tolerable downtime): the total outage of a business process that the system owner is willing to accept.
- RTO (recovery time objective): the longest a resource can stay unavailable before the impact becomes unacceptable. NIST notes that RTO normally has to be shorter than MTD.
- RPO (recovery point objective): the point in time before the outage to which data can be restored from the latest backup. It measures how much data loss is tolerable, not how long the outage lasts.
The same publication separates a disaster recovery plan (DRP), which moves systems to an alternate site, from the contingency plan for a single system (ISCP). Its planning process runs from policy and BIA through preventive controls, recovery strategies and the plan itself to testing, training, exercises and plan maintenance. An ISCP Coordinator runs the process, while the system owner decides how much downtime is acceptable.
Recovery Strategies in the Cloud
AWS architects describe four strategies in a post on aws.amazon.com, ordered from cheapest to fastest:
- Backup and restore: the recovery Region holds only backups, and infrastructure is rebuilt after the event.
- Pilot light: data is replicated continuously, but compute stays switched off until it is needed.
- Warm standby: a scaled-down copy of the environment runs in the recovery Region and has to scale up before failover.
- Multi-site active/active: several Regions serve traffic at once, and failover means routing requests away from the failed one.
The same post on aws.amazon.com stresses that replication is no substitute for backups. If data is deleted or corrupted, the damage replicates too, and the way out is a point-in-time restore to the last known good state. Google Cloud’s planning guide (cloud.google.com) adds the economic rule: the smaller your RTO and RPO, the more the solution costs to run. That is why recovery targets are set per service rather than once for the whole company.
Microsoft’s guidance on designing a disaster recovery strategy (learn.microsoft.com) adds a point worth planning for early: treat failback, the move back to the primary region, as a separate process with its own plan, rather than as failover in reverse.
Backups That Survive Ransomware
The Azure Backup documentation on learn.microsoft.com (a page updated in November 2026) extends the familiar 3-2-1 rule to 3-2-1-1-0: three copies of the data (production plus at least two backups), two storage types or configurations, one copy away from the production site, one copy that is isolated or immutable, and zero restore errors, achieved by monitoring backup jobs and testing recovery regularly.
An immutable backup cannot be deleted, and its retention cannot be shortened, before it expires. In Azure this is the immutable vault. According to learn.microsoft.com, immutability can be switched off again until you lock the vault; once locked, the setting is irreversible, and backups use WORM (write once, read many) storage wherever the service already supports it. In regions and for workloads where WORM is not yet available, the backup service enforces the lock itself, and backups move to WORM storage automatically once it becomes available, with no action needed from you. AWS Backup, according to its service page on aws.amazon.com, also supports immutable backups and logically air-gapped vaults. Names and conditions differ from cloud to cloud, so check the documentation for the specific service.
The same page on learn.microsoft.com is blunt that a successful backup job does not prove a workload can be recovered. The proof is a restore into an isolated test environment, a check that the application works, and a recorded recovery time measured against the RTO.
What NIS2 and DORA Require
The NIS2 Directive (2022/2555, text on eur-lex.europa.eu) lists business continuity among the minimum risk-management measures in Article 21(2)(c), naming backup management, disaster recovery and crisis management.
For the financial sector, the DORA Regulation (2022/2554, eur-lex.europa.eu), which has applied since 17 January 2025, goes into detail. Article 12 requires:
- backup policies that define the scope of the data covered and the minimum backup frequency, based on how critical or confidential the data is;
- restoration and recovery procedures and methods, tested periodically;
- when restoring on their own systems, the use of ICT systems physically and logically segregated from the source system;
- recovery time and recovery point objectives set for each function, taking into account whether it is critical or important.
Our article on DORA, linked below, covers the regulation more broadly.
Read Also
- Ransomware: Preparing Your Company for an Attack: getting ready before an attack
- DORA: Digital Operational Resilience for the Financial Sector: rules for financial firms
- Multi-Cloud Strategy: An Implementation Checklist: running across clouds
Build Your Skills
Designing for resilience and recovery in cloud architectures is part of the Cloud Architect course, while BIA, RTO, RPO and testing continuity plans are covered in Business Continuity Management. Get in touch and we will tailor the programme to your environment.
Frequently Asked Questions (FAQ)
What is business continuity, and how does it relate to disaster recovery?
Business continuity is about keeping business processes running during and after a disruption. Disaster recovery is its technical side: bringing IT systems back, often at another site. NIST SP 800-34 Rev. 1 (csrc.nist.gov) describes the disaster recovery plan as one that supports a broader business continuity plan.
How do you keep a business running in the cloud?
Start with a BIA and set RTO and RPO for each service, then choose a strategy, from plain backups to active/active Regions. Google Cloud’s guide (cloud.google.com) points out that tighter targets cost more, so not every service needs the fastest option. Rehearse the plan regularly.
How should you secure backups in an IaaS cloud?
Keep backups outside the production account and Region, turn on immutability and lock it, and restrict who can delete recovery points. The Azure Backup documentation (learn.microsoft.com) also recommends multi-user authorisation and soft delete. Restore into an isolated environment on a regular schedule.
What do NIS2 and DORA require for backup and recovery?
NIS2 Article 21 makes backup management and disaster recovery part of business continuity. DORA Article 12 requires financial entities to have backup policies, tested recovery procedures, restores on systems segregated from the source, and RTO and RPO for each function. Both texts are on eur-lex.europa.eu.