Imagine a campsite operator whose booking application is hosted by a cloud provider and distributed across several availability zones in the same region. A maintenance script deletes the production database by mistake, and the replicas apply the deletion. No zone has gone down, yet the application is at a standstill: what happens next depends on what the company’s cloud disaster recovery plan provides for.
A cloud disaster recovery plan is the disaster recovery plan for applications hosted by a cloud provider. It says how to bring them back into service within the time and with the data loss the business has accepted. We propose building it in six steps: set the objectives of each application, draw the line of responsibilities with the provider, choose the disaster scenarios, organise backups, choose a failover mode, then test.
A cloud disaster recovery plan rests on objectives per application and on a written boundary with the provider
The two recovery objectives, acceptable interruption time and acceptable data loss, are defined in our entry RTO RPO. The business sets them for each application. In the cloud, the inventory that carries them extends to what the application needs in order to restart: managed services, software as a service, directory, name resolution, infrastructure code repository, vault for secrets and keys. If the repository used to rebuild an application has no objective, the application’s objective does not hold.
Next comes the boundary. Some providers publish its terms; on 12 November 2026 we read the AWS white paper “Disaster Recovery of Workloads on AWS: Recovery in the Cloud”, the Azure reliability pages on Microsoft Learn and Google Cloud’s disaster recovery guide. AWS distinguishes “resiliency of the cloud”, for which it is responsible as regards the infrastructure that runs its services, from “resiliency in the cloud”, which falls to the customer and varies according to the services chosen. For a managed storage or database service, the customer remains responsible for the resilience of its data, backup, versioning and replication included.
Microsoft describes three levels for Azure: platform reliability, for which it alone is responsible; capabilities (availability zones, regions, backups) that it provides and that the customer chooses and configures; the application, which falls to the customer. It writes that this recovery is not an automatic feature of Azure (“DR isn’t an automatic feature of Azure”).
The result is a register with one line per service used: what the provider ensures according to its documentation and the contract, whether or not the resilience option is enabled, the objective of the application served, the person responsible for recovery.
The scenarios that high availability does not cover
Distributing an application across several availability zones is a matter of high availability. At AWS, a region groups several physically isolated zones, which gives better protection against the loss of a data centre. Microsoft distinguishes common risks, which high availability must absorb, from rare risks with major effects, which fall under disaster recovery.
The plan records in writing the events that the architecture does not absorb on its own:
- the loss of an entire region;
- the deletion or corruption of data, which replication copies like everything else;
- the compromise of an administration account, ransomware included;
- the unavailability of the provider’s administration tools at the moment of acting.
Each scenario receives a response, or a written acceptance of the risk by the business. Data location weighs on this choice: Microsoft describes the case where recovery must remain in the original region, with longer timescales and a residual risk to document. The threshold separating a routine incident from a disaster and the person authorised to trigger the plan are written down at the same time. Handling routine incidents is covered by our page SRE and observability.
Backups and failover mode follow from the objectives
Backup responds to the scenarios that replication makes worse. Dated restore points make it possible to go back to before the error or the attack, and their frequency caps the data loss. The copies reside away from production: in another region according to Microsoft, in another account according to AWS, which sees this as protection against the insider threat and account compromise. An immutable backup adds a lock. The scope goes beyond data: AWS points out that infrastructure, configuration and code must also be redeployed, which a description as infrastructure as code makes possible.
The failover mode sets what runs permanently on the recovery site, and therefore the recovery time and the running cost. Google Cloud speaks of “cold”, “warm” and “hot” patterns; AWS describes four strategies, in increasing order of cost and complexity, which the table reproduces without recommending this provider.
| Strategy (AWS term) | Permanently on the recovery site | To do on the day of the disaster |
|---|---|---|
| Backup and restore | The backups, copied away from production | Redeploy infrastructure, configuration and code, then restore the data |
| Pilot light | Replicated data and the infrastructure foundation; application servers ready but switched off | Switch on the servers and scale up |
| Warm standby | A functional copy of production, running at reduced capacity | Scale up |
| Multi-site active/active | The application, running in several regions at once | Redirect traffic; data corruption still calls for a backup |
The strategy is chosen for each application, according to its objectives: one that tolerates a day of downtime need not fund a permanently running standby.
What remains is to write down the failover: who decides it, in what order the components restart, how traffic is redirected, how the return is made. Microsoft points out that this return (failback) is complex to prepare, because of the data written during failover. Google Cloud recommends that each task in the plan be a concrete action, down to the command to run. AWS advises failing over using only the data plane, which delivers the service, without depending on the control plane, which configures the environment.
A timed test, then a variance to address
All three sets of documentation call for the plan to be tested regularly. Microsoft includes human procedures; Google Cloud suggests a test in which the usual access path to the provider does not work.
The test gives two measurements: the recovery time and the date of the most recent data restored. Let us return to the opening scenario, with example values that are not recommendations. The operator has set an interruption time of six hours and a data loss of two hours. It takes an hour and a half to redeploy the infrastructure from code, four hours to restore the database and an hour and a half for checks by users, seven hours in all. The last usable backup is eight hours old. Both objectives are missed, the first by one hour, the second by six hours. The business relaxes its objectives, or funds more closely spaced restore points and a pilot light.
Article 11 of the DORA regulation requires financial entities, outside the simplified framework, to test their response and recovery plans; Article 21 of the NIS2 directive counts business continuity among the measures that each Member State must impose on essential and important entities. Applying them to a given case falls to the organisation’s lawyers. Our page cyber resilience describes how these plans are built with the business and put to the test in crisis exercises.
Each test leaves a dated report: the durations measured, the variance against the objectives, the open actions and the date of the next one.