The meaningful question after a cloud failure is not whether data was copied, but whether a priority workflow can return within its agreed recovery-time and data-loss limits. That requires a recovery design that accounts for dependencies, access and a tested sequence of actions, not merely retained backups.

Consider a familiar cloud incident: a privileged account is compromised, a bad infrastructure change spreads further than expected, and production is no longer trustworthy. The database snapshots may still be present. But can the team reach the recovery environment without using the compromised identity system? Is the required Domain Name System (DNS) access available? Are encryption keys, secrets, container images, network rules and third-party integrations recoverable in the right order? Can anyone say which point in time is safe to restore?

That is the gap between backup and cloud resilience. Backup is a control for preserving recoverable data and, sometimes, configuration. Disaster recovery is the planned restoration of a service after a serious disruption. Resilience is broader again: the technical design and operating capability that lets a service withstand faults where possible, contain their impact, and recover predictably where it cannot. Treating those as synonyms is how organisations discover, at the worst possible moment, that their recovery plan was really an optimistic storage policy.

The practical test: if a team cannot restore a usable customer or business workflow to a stated standard, within a stated time, using credentials and procedures that remain available during the incident, it does not yet have demonstrated resilience. It has backups.

Start with the service that must return, not the backup product

A left-to-right flow from service objectives, through a recovery contract and dependency-aware recovery design, to rehearsed recovery and demonstrated resilience.
From Recovery Objectives to Proven Resilience — Simplified conceptual flow showing how business recovery targets become tested service recovery. Diagram: isageek.

The right starting point is not “how often should we back up?” It is “what must people be able to do again, and by when?” A payroll run, online booking path, case-management workflow and public website may sit in the same cloud environment while needing entirely different recovery investments.

That decision should result in two plain measures. Recovery time objective (RTO) is the maximum tolerable time before the service is restored. Recovery point objective (RPO) is the maximum tolerable period of lost data. Neither is a technical aspiration to pick from a dropdown: both are business tolerances that architecture must either meet or challenge. AWS and Microsoft make the same essential point in their reliability guidance: set recovery targets from business impact, then choose and validate the recovery strategy against them. (docs.aws.amazon.com)

A two-hour RTO and a 15-minute RPO imply very different engineering from a service that can be unavailable until the next working day and restored from the previous night’s copy. The former may justify replicated data, pre-provisioned capacity and rehearsed traffic failover. The latter may sensibly use backup and restore, provided the organisation has timed the full process rather than assuming it will be quick.

This is where broad “cloud-native resilience” claims need some resistance. Running in a public cloud does not automatically create multi-region recovery, application-level failover or protection from an authorised but destructive change. Cloud providers offer highly resilient building blocks; customers still decide the workload architecture, recovery priorities, data model and access controls around them. Microsoft explicitly notes that the customer, not the platform, decides what is mission-critical. (learn.microsoft.com)

Requirements and assumptions worth writing down

Before choosing cold standby, warm standby or anything more elaborate, write a short recovery contract for each important service or workflow. It should name the owner, RTO, RPO, recovery priority, acceptable degraded mode and the decision-maker authorised to declare disaster recovery.

Then document assumptions that normally live only in someone’s head. Does recovery require the primary cloud control plane, the provider interface used to manage resources? Is a second region genuinely suitable for every managed service in use? Can the application operate while a reporting platform, email gateway or payment provider is unavailable? What happens to queued messages and in-flight transactions? Does the recovery design rely on DNS, a content-delivery network, a software-as-a-service (SaaS) identity provider or a network circuit that has not been included in the test?

Dependencies decide whether a restore becomes a service. A database restored perfectly but pointed at an absent message queue is not a recovered application. A web tier brought up before its secrets, certificates and firewall rules are available is not a recovered service either. Azure’s current guidance specifically warns teams to account for dependencies between data stores and to validate all dependencies after recovery, including in partial-failover designs. (learn.microsoft.com)

Question What it changes
How much data can be lost? Backup frequency, point-in-time recovery, replication method and the need to reconcile transactions after recovery.
How long can the service be unavailable? Whether infrastructure can be rebuilt on demand or must already exist in a recovery location.
What has to work first? Recovery sequence: identity, network and secrets often precede applications; applications may precede reporting and batch processing.
What failures are in scope? Whether a separate account, tenant, region, provider-independent copy or manual fallback is required.
Who can authorise and perform recovery? Break-glass access, meaning emergency access for exceptional situations, role separation, credential storage and the design of runbooks.

Design principles: separate failure domains, not just storage locations

A copy in another bucket, vault or disk tier is useful. It is not necessarily independent. If the same compromised administrator can shorten retention, delete copies, alter backup policies or revoke the keys needed to decrypt them, the recovery path shares the same failure domain as production.

The answer is not a magic “air gap” label. It is deliberate separation of authority. Put recovery copies under different administrative control where the risk warrants it; use least-privilege roles, limited to necessary permissions; require a second approval for destructive backup actions; and protect recovery credentials independently of ordinary production administration. Cloud platforms support versions of this pattern. AWS recommends copying backups into a different account as an additional barrier against compromise of one account, while Azure’s Resource Guard model can separate authorisation for sensitive vault operations from normal backup administration. (docs.aws.amazon.com)

For the most important workloads, test the uncomfortable question: “What can an attacker with a tenant administrator or subscription owner role do?” If the realistic answer is “delete the backups and disable the controls that protect them”, the arrangement is not sufficiently isolated.

Immutability matters here, but it is a retention control rather than a complete recovery strategy. It can stop backup data being altered or removed during its protected period; it cannot provide a clean application configuration, decide the correct recovery point, restore a dependent SaaS service, or ensure that the people running the incident have usable access. AWS also warns that compliance-style vault locks can create persistent retention costs if policies are set carelessly. That is a reminder that resilience controls need lifecycle ownership as well as security enthusiasm. (docs.aws.amazon.com) For the limits of the control itself, see what immutable backup protects and what it does not.

Choose the recovery architecture from the objective

There is no honour in active-active architecture for a back-office system that can safely be restored tomorrow. Equally, there is no credibility in promising near-continuous service from a design that requires rebuilding networks, restoring terabytes of data and discovering the right application version during an outage.

At one end, backup and restore is usually the cheapest and simplest design. Infrastructure can be recreated from code, applications redeployed, then data restored to a recovery location. It is often entirely appropriate for lower-priority services, but its RTO includes every slow step: acquiring capacity, restoring data, applying configuration, checking integrity and returning traffic.

A pilot light keeps essential data and a small core of the environment ready, scaling the rest during recovery. Warm standby keeps a reduced but working service running elsewhere, allowing faster recovery at recurring cost. Active-active can reduce interruption still further, but it introduces the difficult work of multi-region traffic routing, data consistency and avoiding replication of bad changes. AWS positions these approaches as increasing in cost and complexity while generally reducing recovery times and data-loss exposure. It also makes an important qualification: replication alone does not protect against corruption or destructive change without point-in-time recovery or versioned data. (docs.aws.amazon.com)

The sensible design is often mixed. Keep the customer-facing transaction path on a warm or active design where the impact supports it; recover reporting, archives and internal tools later from protected copies. More importantly, divide large services into recoverable components where possible. A contained failure and partial failover, moving only the affected component to a recovery environment, are usually cheaper and less risky than declaring the entire estate a disaster.

Operational considerations: recovery access is part of the architecture

Cloud recovery plans commonly underestimate identity. If single sign-on, privileged access, secrets management or administrator multi-factor authentication (MFA) depends on the damaged environment, the technical recovery path may be inaccessible. Maintain tightly controlled emergency access accounts, document how they are used, and ensure the necessary people can reach the recovery environment from approved devices and networks.

That does not mean creating permanent all-powerful accounts and hoping nobody finds them. It means a governed break-glass process, separately protected credentials, monitoring and a test of the whole sequence. CISA’s cloud guidance similarly recommends considering emergency global-administrator accounts for situations where normal cloud access is unavailable. (cisa.gov)

Secrets, certificates, encryption keys, infrastructure-as-code state, image registries and software artefacts need the same scrutiny. A team that has backed up a database but cannot retrieve the application’s keys has not met its RTO. Nor has a team whose Terraform or equivalent automation refers to deleted accounts, unavailable modules or a control-plane service that is part of the incident.

Identity deserves its own recovery treatment in a cyber incident. The practical issues around email, OAuth, an authorisation framework, consent and privileged accounts are covered in ransomware recovery that starts with email, identity and OAuth; the point here is narrower but crucial: no cloud recovery design is complete if its operators cannot safely enter it.

Test the outcome, not the backup dashboard

A green backup dashboard proves that scheduled jobs completed according to the tool. It does not prove recoverability. Periodic restoration is an explicit AWS backup recommendation, and Google’s reliability guidance calls for recovery tests that include regional failover, release rollback and restoring data from backups. (docs.aws.amazon.com)

Testing should progress from simple to realistic. Restore a representative data set and verify it. Rebuild a service from infrastructure code in a separate environment. Exercise the documented recovery order. Fail over a component. Finally, for services whose downtime matters, run a controlled scenario in which people, identities, integrations, traffic routing and communications all participate.

Measure the result against the agreed target. Record the actual recovery point, elapsed restoration time, manual interventions, unavailable dependencies and data-reconciliation work. If the measured RTO is eight hours, the service does not have a two-hour RTO simply because that number appears in a policy document.

Tests also expose configuration drift: the steady accumulation of changes that make a recovery environment unlike production. Automating infrastructure and deployment reduces that risk, but does not abolish it. Runbooks must have owners, changes must trigger recovery-plan review, and findings must change either the design or the stated objective. NIST’s contingency-planning guidance remains useful on this point: recovery is a coordinated set of plans, procedures and technical measures, not a backup task delegated to storage. (csrc.nist.gov)

Decision points for a credible cloud resilience plan

  • Define service tiers from impact. Start with business workflows, not application inventories. Set RTO and RPO for each tier and identify what may operate in a reduced mode.
  • Map dependencies and recovery order. Include identity, network, DNS, certificates, keys, SaaS integrations, code artefacts and external suppliers. Make hidden dependencies visible before an incident does.
  • Protect copies and recovery authority separately. Use retention controls, distinct administrative boundaries and approval mechanisms appropriate to the threat model. Never assume a backup is safe merely because it is in another region.
  • Use infrastructure as code, but retain a usable recovery path. Version it, protect its state and verify that it can build the required environment under incident conditions.
  • Practise the failure modes that matter. Include accidental deletion, malicious or mistaken configuration change, loss of a region, loss of control-plane access and compromise of privileged credentials.
  • Publish evidence, not reassurance. Report the latest recovery test date, achieved RTO and RPO, scope, exceptions and corrective actions to the people accepting the risk.

For most organisations, the immediate improvement is not an expensive multi-region rebuild. It is a disciplined recovery design for the services that matter most: written objectives, dependency maps, independent protection for recovery copies and credentials, executable runbooks, and tests that demonstrate a usable service rather than a completed restore job.

That approach also prevents waste. Some workloads should be restored slowly because their impact is low. Others deserve warm capacity, replicated data or carefully designed active operation. The decision is not whether cloud resilience is good. It is whether the resilience being bought, built and operated matches the consequence of failure.

Backup remains essential. But it is only one of the ingredients. A resilient cloud service is one that can be recovered in the real conditions of an incident: with the right data, in the right order, by people who can still gain access, within a limit the organisation has consciously chosen. For a broader distinction between copies and recovery planning, a backup that is not yet a disaster-recovery strategy is the right starting point. Cost modelling should come afterwards, once the required recovery posture is clear; the practical trade-offs are explored in how much Azure disaster recovery really costs.

Spot an error?

If something factual looks wrong, outdated or misleading, flag it here. Corrections are reviewed separately from normal article comments and reader questions.