Roll Back After Migrating to OpenShift Virtualization
Plan rollback from OpenShift Virtualization to VMware before cutover. Control production writes, preserve the source, and rehearse how changed data returns.
Personal site note: views expressed here are my own and do not necessarily represent Red Hat or any other organization. The migration controls below refer to Migration Toolkit for Virtualization (MTV) 2.12. Rehearse the return procedure with the application and database versions you actually run.
The virtual machine boots on OpenShift Virtualization. The application owner signs in. Someone on the change bridge says the VMware copy is still there, so rollback is covered.
Then the application accepts an order.
If that order exists only on the new VM, powering on the old one brings back an earlier version of the business. The source disks can be intact and the rollback plan can still be wrong.
For a VMware vSphere migration, the useful question is what has changed since the source stopped, and how will those changes survive a return? Before the target changes production state, returning may be a controlled restart of the source. Afterward, it needs a data recovery or application failback procedure as well.
The Oracle migration post uses a database standby to keep that return path current. A general virtual machine (VM) migration needs the same conversation, even when the workload is a file server, a queue consumer, or a small application with its database on the boot disk.
What MTV completes
Migration Toolkit for Virtualization (MTV) moves VMs to OpenShift Virtualization. With warm migration, disk copying starts while the source runs. Cutover shuts the source down and transfers the remaining data. Cold migration copies with the source shut down. Neither approach carries the running guest’s memory into the new VM.
On the default path, MTV uses virt-v2v to prepare the copied guest for the destination, including changes to drivers and boot configuration. Raw copy mode skips that conversion. Treat a later return of the converted disks to VMware as another compatibility and migration exercise. Do not put an untested disk export in the emergency procedure.
MTV also lets you cancel an in-progress migration. The documented behavior leaves already-migrated VMs unchanged. Canceling the rest of a wave does not return its completed VMs to vSphere or move their new application data back.
Record migration completion and application acceptance separately. The first says the move finished. The second says the service works and its recovery procedure is ready for the next failure.
Find every place the application can write
Start with the application owner and follow one transaction. Record the database, uploaded files, shared storage, message queues, and outside services it touches. Include jobs that run without a user signing in.
A moved web server with a database that stays in place may have little local data to return. It still needs an application version compatible with the current database schema. A moved database VM has a different problem: its old disks stop receiving transactions when the source stops.
The first important write might happen before you change DNS. A service can start a scheduled job, consume a message, rotate a shared credential, or call a payment API as soon as the guest boots. Removing the VM from a load balancer only controls the traffic that passes through that load balancer.
An isolated boot also writes logs and operating-system state. Decide which changes are disposable and which affect production. The boundary is the first change the business needs to preserve, whether it lands on the VM or somewhere else.
For each application, write down three things:
- Where authoritative data lives. Include dependencies outside the migration wave and the owner who can verify their state.
- What can change it. User requests, integrations, scheduled tasks, queue consumers, and startup actions all count.
- How it returns. Name the replication, restore, replay, or reconciliation procedure, and who has rehearsed it.
If the third answer is only “the old VM is still powered off,” the procedure covers an unchanged application state. Keep production writers closed until the plan addresses the next state too.
Walk the cutover clock
Consider a warm migration of an order application and its database. These times are illustrative; the checkpoints belong in the change record.
- 21:50 — Close production writers. Put the application into its maintenance state, drain active work, stop consumers and scheduled jobs, and close the database cleanly. Verify the application is quiet before the final transfer. Record the last accepted transaction or another application checkpoint the owner can verify.
- 22:00 — Cut over the VMs. Complete the final transfer. Confirm the source VMs are off and protected against an accidental restart. Keep their disks, configuration, and required vSphere capacity available.
- 22:20 — Validate the target with production access blocked. Check boot, disks, network paths, application startup, and the recorded data checkpoint. Keep background writers and outbound integrations under the same controls as user traffic. Any test writes must have an agreed disposition.
- 22:35 — Authorize production writes. The application owner accepts the checks and the named operator opens the service. Record this time separately from MTV completion. From here, use the procedure that preserves new transactions if you need to return.
- 23:05 — A fault appears under load. Stop admitting new work and identify the authoritative copy. If the target accepted orders, those orders must survive the recovery decision. Booting the old database cannot recover records its disks never received.
At 22:20, a failed check may still lead to a straightforward source restart. At 23:05, the same symptom can require replication, restore, or transaction reconciliation. Elapsed time alone does not choose the procedure. Changed application state does.
Make the validation window real
MTV 2.12 exposes a target power-state setting: off, on, or auto. The auto setting preserves the source power state. For a wave that requires an explicit startup decision, select Powered off and verify that behavior in the rehearsal.
That setting gives the operator control over startup after migration. It does not define which destinations the guest can reach once started. Prepare the validation network and application controls before booting the target, and review migration hooks for actions that change external systems.
Use the actual VM network attachment to decide where isolation is enforced. Have the network owner prove that the target cannot reach prohibited production endpoints. A namespace label or an untested policy is insufficient evidence for an application attached to an existing VLAN.
Keep duplicate hostnames and addresses separated. Give testers only the access the rehearsal requires. If authentication or another shared service must remain reachable, include its possible state changes in the return procedure.
Test writes need the same attention. An order that sends an email, reserves inventory, or reaches a payment service leaves effects outside the VM. Deleting its local database row does not undo those effects. Use test endpoints or a documented reconciliation procedure.
Return before production state changes
For a retained source with no production changes to preserve, the return sequence is relatively small. Rehearse it for the whole application:
- Stop any remaining migration or automation activity that could start a target or change its routing. Confirm the resulting state.
- Stop and isolate the target VMs. Verify they cannot serve traffic or resume background work while the source returns.
- Check dependencies outside the VMs for changes made during validation, including credentials, application configuration, and test activity.
- Start the retained source VMs in the application dependency order. Verify their data checkpoint and service health.
- Restore the intended client path and reopen writers. Check actual client connections; DNS caches and established sessions can outlive a routing change.
Assign one person control of the source/target handoff. A powered-off VM can still be started by another administrator or an automation job. Record the controls that prevent that, and verify them on both platforms.
Keep the target disks and logs available for diagnosis until the application owner releases them. Recovery is a poor time to discover that cleanup removed the evidence needed to explain the failure.
Return after the target accepts writes
Once production state changes, choose the recovery path based on where the current data is and how it can be preserved.
Repair on OpenShift. A network mapping or application configuration problem may be faster to correct in place while retaining the current data. Compare that work with the measured return procedure before committing to another move.
Switch back through application replication. If the old environment has remained a valid replica, use the application’s role-transition procedure. Oracle Data Guard, for example, supports a planned switchover between a primary and standby without data loss. The databases continue in their new roles. Verify synchronization and readiness before a switchback, and include application connections in the rehearsal.
This is why the Oracle design keeps the former primary participating as a standby. That VM is intentionally running in a replication role. The power-off rule for an independent, stale source copy does not apply to a correctly managed standby.
Recover current data into the return environment. Without a replica, the application may need a native backup and restore, log recovery, or an export and import into a compatible guest. Measure that full path, including validation. Reconcile external messages and transactions that a database restore cannot reverse.
If current data cannot be recovered within the outage budget, the business owner has a decision about outage duration and possible data loss. Record exactly what would be lost before taking that path. A healthy old VM does not make the decision disappear.
The backup and disaster recovery distinction still applies. A restore point has a time and a scope. Identify the latest recoverable transaction and the destination where the restore has actually been tested.
Budget time for the return
Put the latest decision time in the change record. Calculate it from the service restoration deadline, the measured recovery duration for the current application state, and a margin for validation.
For example, if service must be available by 02:00, a rehearsed return takes 45 minutes, and the agreed margin is 15 minutes, the decision is due at 01:00. Those numbers are a planning example, not an MTV estimate. If new writes require a two-hour restore, the 45-minute source-restart measurement no longer applies.
Name the person who can stop the attempt, the application owner who accepts recovery, and the operators for both platforms. Put observable failure conditions beside the deadline: a required transaction fails, data checks disagree, or an essential dependency remains unavailable.
Rehearse two failures in non-production. First, reject the target before production access opens and prove the source returns. Then commit identifiable test transactions on the target, invoke the recovery path, and prove those transactions survive. Include a queue message or an external test integration if the application uses one.
Retire the source after recovery is proven
Keep source capacity through an agreed acceptance period that includes the important business jobs. A successful login says little about the overnight batch or the next scheduled export.
Before retiring the source, require application acceptance, functioning monitoring, and a successful backup and representative restore from the new environment. Record the recovery procedure that will apply after the old VM is gone.
Preserve the migration evidence before housekeeping. MTV’s archive operation removes plan logs, history, and metadata and is irreversible. Export what the operations team needs before archiving the plan.
Retaining an unreplicated source for seven days does not give it seven days of new transactions. Retention buys time to investigate and keeps an environment available. A current return path still needs current data.
The solutions architect takeaway
- Record the write boundary. Application startup, background jobs, and external integrations can cross it before user traffic opens.
- Keep a controlled validation window. Choose the target power state, prove isolation, and define how test changes are handled.
- Write two return procedures. One for an unchanged production state, another for preserving transactions accepted after cutover.
- Measure the whole recovery. Include data movement, dependency checks, client routing, and application acceptance in the deadline.
- Prove recovery before cleanup. Preserve the source and migration evidence until the new service and its recovery path are accepted.
For the next rehearsal, migrate one non-production application, write a transaction on the destination, and recover it through the proposed return path. The transaction is the proof that matters to the person who owns the service.
Related posts
- Migrate Oracle Databases to OpenShift Virtualization
- vSphere to OpenShift Virtualization: Translation Guide
- OADP for OpenShift VMs: Backup Is Not DR
- AI Agents for MTV: vSphere to OpenShift Virtualization
Want help turning a migration wave into a cutover and recovery plan? Reach out to your Red Hat account team—or rehearse one application return with a transaction that must survive before scheduling production.