← Blog
An open lavender door with charcoal outlines and an orange handle on a pale lavender background.
Published on
2026-11-13

Cloud exit strategy: prove you can leave before you need to

A cloud exit strategy records how a business could leave a provider, what would have to change, and who can decide whether the move is worth doing. It gives a future decision something better than a diagram and a promise that the data is exportable.

Start with one important application. Identify the services it depends on, collect evidence that its data and software can work elsewhere, estimate the cost of replacing the provider’s operational work, and set a review date. The outcome may be a funded migration, a smaller change, or a documented decision to stay.

What should a cloud exit strategy contain?

Keep a short exit-readiness record for each application whose loss or prolonged disruption would matter to the business. Give it an accountable owner and connect it to the application’s technical documentation. A useful record answers six questions:

  • What change would make us reconsider this provider?
  • Which data, application components, and operating responsibilities are in scope?
  • What evidence shows that a replacement could meet our requirements?
  • What remains untested, contractually unclear, or dependent on specialist skills?
  • What would the transition and subsequent operation cost?
  • Who accepts the remaining dependency, and when will they review it?

The Australian Government’s guidance on lock-in and portability offers a useful principle: provider dependencies can be deliberate choices when their value, risks, owners, and review points are understood. Its agency guidance is a reference here, rather than a requirement imposed on every business.

Once a move is approved, use a separate cloud migration checklist for the delivery work: sequencing, cutover, validation, and rollback. Exit readiness should exist before that approval.

Set triggers that lead to a decision

“Costs are too high” is difficult to act on. A more useful trigger specifies the application, the measure, and the person who reviews it. For example, a fictional customer portal might require a review if its cost per active customer exceeds an internally agreed limit for three consecutive months.

Other triggers could include a contract renewal, a required region becoming unavailable, a service being retired, or a new requirement the current architecture cannot meet. Have the relevant specialist assess legal or contractual changes. A trigger starts an assessment; it does not authorize a move automatically.

Separate a planned commercial exit from an emergency. An export procedure that requires the current provider’s console, identity service, and network may be unavailable during an outage. Record which access paths and recovery copies remain usable in that situation. Keep continuity and disaster-recovery procedures alongside the exit record.

Work through one application, not the entire cloud account

Consider a fictional supplier portal with a web application, PostgreSQL database, uploaded documents, background jobs, and company sign-in. The following is an example of questions to resolve, not a report of completed tests.

DependencyEvidence to collectOwner
Web application and jobsBuild and start from documented source and configuration; list provider-specific APIs and scheduler behavior.Application lead
Managed PostgreSQLRestore a representative export; verify extensions, roles, permissions, and application behavior on the proposed target.Database owner
Document storageCopy representative objects and required metadata; verify access, links, retention needs, and application reads.Storage owner
Identity, secrets, and encryptionShow how users authenticate and services obtain credentials; establish a supported way to access encrypted data after the move.Security lead
OperationsAssign patching, monitoring, incident response, backup, and restoration responsibilities at the destination.Operations lead

A container image addresses only part of this list. The application may still depend on a proprietary queue, identity policy, network route, or deployment service. Likewise, an S3-compatible interface is a starting point for checking object operations, not proof that every permission, lifecycle rule, or event integration transfers unchanged.

For the database, PostgreSQL documents that pg_dump exports a single database; cluster-wide objects such as roles need separate handling. Treat a successful export as one piece of evidence. Restoring it and running the application reveals a different set of dependencies.

Make the evidence record honest

Use three plain statuses: documented, demonstrated, and unresolved. “Demonstrated” should identify the test date, software versions, dataset, destination, result, and limitations. A small sample restore does not establish how long the full production dataset will take.

For the fictional portal, an annual review could produce the following record. These entries illustrate the form; they are not measurements or claims about a Hivenet deployment.

Readiness itemIllustrative recordNext action
Database portabilityExport procedure documented; target restore unresolved.Database owner schedules a representative restore and records compatibility gaps.
File accessCopy method documented; authorization and retention behavior unresolved.Storage and security owners test those behaviors together.
Operating modelApplication team can deploy; overnight incident coverage unassigned.Operations lead compares support options before approving a target.
Commercial exitNotice period recorded; assistance scope and transfer charges awaiting written confirmation.Procurement obtains a dated response for the relevant services.
DecisionStay pending evidence, with a named decision owner.Review before renewal or after a material service change.

Keep test outputs and commercial evidence accessible to the people who would execute the exit. Avoid making the only copy dependent on the account you might lose access to. Apply the organization’s access and retention rules to these records.

Estimate transition cost and elapsed time separately

Price the replacement workload and the work needed to reach it. A lower VM rate cannot tell you whether replacing a managed service is economical. Use the existing cloud cost optimization review to establish the current workload’s baseline, then add the costs specific to leaving.

  • Engineering: export, adaptation, infrastructure setup, testing, and documentation.
  • Data movement: retrieval, transfer, requests, temporary copies, and any required tooling.
  • Overlap: two environments, support arrangements, licenses, and commitments during transition.
  • Operations: additional staff time, training, monitoring, backup, and incident coverage.
  • Closure: contract obligations, evidence retention, access removal, and agreed deletion work.

Record a range and its assumptions. As a planning example only, a team might allocate 4–8 engineer-weeks and 2–4 weeks of overlapping hosting, then leave transfer charges unresolved until it has measured the dataset and confirmed terms. These are placeholders to replace, not industry averages. Engineer-weeks measure effort; procurement waits, staffing, and technical dependencies determine elapsed time.

The NHS cloud exit strategy explicitly considers parallel hosting and the time and cost of data movement. Those are useful budget checks even when the organization and workload are much smaller.

Ask the current provider about export formats, assistance, access after notice, outstanding commitments, and any transfer-fee program. For example, Google Cloud’s portability procedures describe service-specific methods and an exit program with conditions and notice steps. Do not assume a fee waiver applies to every transfer or eliminates migration work. Confirm the terms that apply to your account.

Choose among staying, moving part, and leaving

A useful review compares feasible options. Staying may preserve managed capabilities that the team cannot replace economically. Moving one independent workload may reduce a specific dependency without forcing a platform-wide rewrite. A complete exit may be justified when the provider can no longer meet a critical requirement.

Moving workloads to infrastructure the business operates itself, sometimes called cloud repatriation, is another destination choice. Include capacity procurement, maintenance, staffing, and recovery in that comparison. Ownership of hardware does not remove operating responsibilities.

Write the decision in a form the next reviewer can challenge: “Stay through the next review because the managed service’s operational value exceeds the expected benefit of moving; fund the unresolved restore test before renewal.” Attach the evidence and identify who accepts the remaining risks.

Review the record when the application changes

Choose a review interval that matches the workload’s importance. An annual review is a reasonable example for the portal, with earlier checks before renewal or after major changes. A new database extension, encryption dependency, dataset increase, or loss of key staff can invalidate an earlier estimate.

Track the last demonstrated result separately from the latest documentation update. Refreshing the date on a plan does not refresh its evidence. Give every unresolved item an owner, a next action, and a decision deadline.

Evaluate a defined workload with Hivenet

If the record identifies a portable application or storage workload, Compute with Hivenet offers CPU and GPU instances for software you operate, while Hivenet Object Storage provides a business S3-compatible storage route. Check the required operations, region, capacity, support, and recovery behavior for the specific workload. Do not assume a managed database or complete AWS feature equivalence.

Discuss one workload with Hivenet, bringing its dependency record, dataset size, access pattern, operating responsibilities, and unresolved requirements. Use the conversation to define a bounded evaluation and the evidence needed for a decision.

Your next workload belongs on Hivenet.

Pick one AI, compute, or storage workload and see the difference for yourself. Spin it up in minutes, or let our team map your fastest path to production.