A POC starts with a decision to inform: which workloads can migrate, at what service level and under which operating conditions? Booting a VM answers only part of that question.
The scope must represent the project's difficulties: a critical application, busy storage, network dependencies, backup and the on-call team. Selecting only the easiest VMs produces little evidence for production planning.
Set criteria before testing
Agree on objectives with application owners and operations. RTO expresses the maximum acceptable interruption; RPO expresses acceptable data loss. Measure them at service level with observable start and end points.
Illustrative example: if the agreed objective is recovery in under 15 minutes and the test takes 22, keep the gap visible in the decision. Changing the threshold after testing requires a new business decision.
For each criterion, assign an owner, a measurement method and a consequence for failure. If an essential test cannot run during the POC, mark it as unvalidated and identify the outstanding evidence.
Prepare an acceptance test matrix
Adapt this starting point to the evaluated scope. Run failure tests on test data in an isolated environment, with a plan to return to normal operation.
| Area | Test | Criterion to agree | Evidence to attach |
|---|---|---|---|
| Migration | Transfer a representative VM | Accepted downtime and functional checks | Timeline and application validation |
| Maintenance | Move a VM under load | Interruption and errors within agreed thresholds | Before, during and after measurements |
| HA | Lose a qualification host | Accepted service RTO and remaining capacity | Logs and control transaction |
| Storage | Disk or host failure | Accessible data and acceptable latency | Protection state and reconstruction |
| Backup | Full isolated restore | Accepted RPO, RTO and application consistency | Restore report |
| Rollback | Simulated incident after cutover | Validated procedure and treatment of writes | Runbook execution and checks |
| Operations | Intervention by the target team | Achievable diagnosis and escalation | Observations and training needs |
The outcome must support decisions by workload family. One global “POC passed” indicator conceals differences between applications.
Test rollback before the production wave
Define the point until which rollback is possible, who decides and which data will be retained. A VMware VM remaining available is insufficient if new writes have occurred on Proxmox.
The qualification scenario can follow these steps:
- Back up the initial state and verify the restore mechanism.
- Transfer a test VM using the chosen method.
- Generate identifiable application writes on the target.
- Simulate a reason to roll back and execute the decision.
- Apply the planned data strategy, then check the service.
- Measure the full duration and document any accepted loss.
Application-level replication, backup restoration or another mechanism may be needed. Name it in the runbook. A snapshot does not replace an independent backup or automatically transfer writes back to the previous platform.
Separate maintenance, HA and disaster recovery
Maintaining a working host and suffering a sudden failure require different tests. vMotion and Live Migration covers the first; the VMware HA and Proxmox HA episode examines restart after failure.
For HA, verify protected VMs, quorum, fencing, disk access and recovery node capacity. Measure through to application service recovery.
A local host failure test does not validate site loss. If disaster recovery is part of the decision, include the other site's dependencies, backup availability, networking, identity, keys and return to the primary site. If this remains outside the POC, record it as an open condition before production.
Restore with the chosen backup chain
Use the versions and components intended for production. If Veeam is selected, check the configuration against its Proxmox support matrix and limitations.
With PBS, include integrity verification followed by restoration. The PBS maintenance documentation describes backup verification. This does not replace functional checks of the restored application.
In either case, test access to required credentials and keys from the recovery environment. Retain a business-level check: opening a database, reading a dataset and running a test transaction. A booting guest does not establish that consistency.
Measure performance during degradation
Establish a representative workload baseline, then repeat measurements during maintenance and storage reconstruction. Record application latency, errors, resource use and time to restore the required protection.
If Ceph is considered, our Ceph and SAN comparison helps assess storage options. Include reconstruction load alongside backup and other operations that may overlap.
POC duration depends on cycles that need observation: peaks, batch processing, retention and maintenance. A fixed duration does not guarantee coverage. Record unobserved cycles among the limitations.
Involve the team that will operate the target
Give the team concrete tasks: handle an alert, open a console, diagnose storage access, restore and execute maintenance. Note steps requiring outside assistance.
Assess permissions, logs, escalation contacts and support availability at required hours. Training and procedure handover belong in the project scope.
Deliver a usable decision file
The exit file should contain the tested configuration, results by criterion, evidence, gaps and tests not performed. Add accepted risks, corrective actions, their owners and retest conditions.
A conditional go must name what still blocks production and who will authorise closure. An unvalidated workload stays outside the authorised wave.
Our Proxmox POC engagement follows this validation approach before migration waves. To prepare your scope, request a POC assessment using your recovery and restore criteria.
Sources and scope
Official sources reviewed on 15 October 2026. The matrix and rollback scenario are proposed methods. They do not describe an executed test campaign or a customer result.