Unit 7.4: Migration and Hybrid Cutover
Introduction
FinCorp's on-premises VMware estate holds several hundred virtual machines: core-banking systems, the payment switch, internal tools, the portal that Modules 3 to 6 already rebuilt, and the databases behind them. The board wants the data center closed within two years, and the risk office wants every outage announced in advance and every step reversible. A migration of that size fails on the plan long before it fails on the tooling.
The platform brings VMs in through three engineered paths: disk images converted on upload, VMware-native replication into a dedicated VMware environment, and machine-level backup and restore. This design unit chooses a path per workload, groups workloads into waves, and attaches a downtime, a gate and a rollback to each wave. It is the plan the whole course has been building toward.
1. Disposition Before Path
1.1 Three Dispositions
Every workload gets one disposition before anyone chooses a tool:
| Disposition | What changes | FinCorp candidates | IONOS CLOUD target |
|---|---|---|---|
| Rehost on VMware | Nothing in the guest, the VMware operating model stays | Core-banking systems, vendor appliances certified on vSphere | Private Cloud, through VCDA (Section 2.2) |
| Rehost on KVM | The hypervisor; drivers and boot firmware in the guest | Payment switch servers, internal tools | VDC compute, through image conversion (Section 2.1) or backup and restore (Section 2.3) |
| Replatform | The platform under the application | Databases moving to DBaaS; file shares to Network File Storage or Object Storage | Managed services, through dump and restore (Section 4) |
| Refactor | The application's architecture | The portal and partner API, already stateless and containerised | VDC compute, VM Auto Scaling, Managed Kubernetes (Modules 4 and 6) |
The disposition decides the risk, not the size. A rehost on VMware changes the least in the guest and moves the most risk into the network and the cutover. A replatform changes the most in the stack and the least in the network. Retiring a workload is the cheapest disposition of all; FinCorp's inventory review removes about one VM in eight before planning starts.
1.2 Inputs Per Workload
The inventory records, per VM: its dependency cluster (Section 3), its RTO and RPO from Unit 7.1, its boot firmware (BIOS or UEFI), its guest operating system and driver state, its disk sizes, and whether its IP addresses are referenced elsewhere. Boot firmware matters more than it seems: the Intel Sierra Forest and AMD EPYC Turin CPU families require UEFI for booting, so a BIOS guest can only land on a CPU family that still supports it, or be converted.
2. The Three Infrastructure Paths
2.1 Path 1: Image Conversion and Upload
The guest's disks become images on the VDC platform. Supported HDD formats are vmdk (VMware), vhd and vhdx (HyperV), cow, qcow and qcow2 (Qemu), raw (binary), vpc (VirtualPC) and vdi (VirtualBox), so a VMware export is taken as its vmdk disks. You upload them to the FTP server of the target location, for example ftps://ftp-fra.ionos.com for Frankfurt or ftps://ftp-txl.ionos.com for Berlin, over FTP with explicit TLS on port 21. Images are only available in the region where they were uploaded, so a workload that must run in two regions is uploaded twice.
Three facts shape the timing and the cost. The image is converted to RAW after upload, so a dynamic disk is always used at its maximum size: a 3 GB file from a 50 GB disk becomes a 50 GB image. Conversion takes a few minutes depending on size, after which the image appears in the image list. And after conversion, the image's metadata decides how it behaves: set OS Type, the Live Vertical Scaling properties and UEFI Support in the image manager before you create volumes from it.
Guest preparation is the engineering part:
- Windows needs VirtIO drivers to recognise the VirtIO SCSI controller and network adapter of the KVM-based hypervisor. Install them on the source before export, or after import by setting the storage bus to IDE temporarily and booting the Windows VirtIO drivers ISO, available under Microsoft Images in the CD-ROM selection. Without them, Windows usually stops with a blue screen at boot. Switch back to VirtIO afterwards; IDE is for troubleshooting only (Unit 7.3).
- UEFI guests may need their boot path. Some operating systems register exact boot loader paths in the original VM's NVRAM. A new server starts with empty NVRAM and finds the loader only on the standard fallback path
\EFI\BOOT\BOOTX64.EFI. Configure the fallback before export. - Network configuration changes. The guest gets new network cards on the platform; addresses come from the VDC's LANs and the address plan of Unit 3.1.
Downtime is the whole copy: the source must be stopped for a consistent disk, so the outage runs from shutdown through export, upload, conversion, provisioning and first boot. Use this path for VMs whose disks are small enough for that window, or whose service can be drained first.
2.2 Path 2: VMware-Native Replication into Private Cloud
For workloads that keep their VMware operating model, VMware Cloud Director Availability (VCDA) moves them into Private Cloud (Unit 4.4). VCDA is a Disaster Recovery-as-a-Service solution that protects vApps and virtual machines with asynchronous replication, migrates them, performs failover and reverses failovers.
The mechanics are specific:
- Two appliances. The appliance in the Private Cloud (VCDA-CLOUD) is installed and configured automatically once you activate the option; its only parameter is the public IP address through which the on-premises appliance reaches it. The on-premises appliance (VCDA-CUST) is deployed from an OVF template in your vCenter and paired with the cloud site under Peer Sites.
- Ports. The on-premises site connects outbound on TCP 443; the cloud site uses TCP 55443, on the main IP address of the Private Cloud. Internal paths at each site add further ports, such as 8048, 902 and 44046, which belong in the firewall change request.
- Operation. Once paired, VMs are migrated or protected from the vSphere Client context menu, the vSphere plug-in, or the VCDA-CUST interface. The documented version is 4.7.x.
- Cost. Migration can be used free of charge; only protected VMs are invoiced, at 50,00 € per VM per month.
The guest does not change: no driver injection, no image conversion. Because replication runs continuously in the background, the cutover is short: stop the VM at the source, let the final changes replicate, and start the replica at the cloud site. Downtime per VM is that final synchronisation and the boot, not the full copy of Path 1.
The network still changes. The Private Cloud's NSX segments are virtual Layer 2 domains that can be linked to specific IP address ranges, so the migrated VMs can keep their addressing scheme, including Bring Your Own IP (BYOIP) ranges, provided each segment's subnet is more specific than the BYOIP range. During the transition, the same range must not be live on both sides of a routed link at once, so each dependency cluster moves its addresses as one step, and the routing change is part of the cutover runbook.
2.3 Path 3: Backup and Restore
The Backup Service agent protects machines inside and outside IONOS CLOUD, including VMware and Hyper-V environments, which makes it a migration path for VMs that land on VDC compute. The flow is a bare-metal recovery:
- Install the Acronis agent on the source VM and run a full machine backup to a Backup Unit.
- Provision a target server in the DCD with a blank volume, connected to its LAN.
- Recover the full machine onto it, either by booting the Acronis Rescue Media ISO through the KVM console, or by installing a temporary operating system and triggering the recovery from the Acronis web console.
- Let the recovery adapt the guest: Acronis adapts the boot partition for a BIOS-to-UEFI change, and for Windows, Universal Restore injects the VirtIO storage drivers.
This is the path for the hard cases: crossing boot firmware or CPU generations, and resizing file systems on the way. The agent needs an internet connection and its ports open to the backup system, which is a firewall change on the source network. Downtime runs from the last backup to the first boot of the restored server; every change on the source after that backup is lost unless the source is frozen before it.
2.4 Comparing the Paths
| Criterion | Image conversion | VCDA into Private Cloud | Backup and restore |
|---|---|---|---|
| Target | VDC compute | Private Cloud | VDC compute |
| Guest changes | VirtIO, UEFI boot path, network cards | None | Adapted by the recovery (boot partition, VirtIO) |
| Data transfer | Full disk at upload, region-bound images | Continuous asynchronous replication | Full machine backup to a Backup Unit |
| Downtime per VM | Shutdown to first boot of the converted image | Final synchronisation and boot | Last backup or freeze to first boot |
| Rollback | Restart the untouched source | Reverse failover | Restart the untouched source |
| Tool cost | None | Migration free; protection 50,00 € per VM per month | Backup Service usage and backup storage |
| Best for | Small, stateless or drainable VMs | Large VMware-bound estates, short windows | Firmware or CPU changes, Windows estates |
3. Wave Planning
3.1 Waves Follow Dependencies, Not Headcount
A wave is a dependency cluster: an application, its database, its batch jobs and the integrations that must answer within its latency budget. Splitting a cluster across the hybrid link turns every internal call into a wide-area call, and the cutover then has two moments of risk instead of one. FinCorp builds the clusters from the flow logs and the application owners' interface lists, then orders the waves by risk: the least critical first, so that the runbook is tested before it meets core banking.
3.2 Links During the Cutover
Each wave needs a link with the right capacity. A VPN Gateway tunnel carries 1 Gbps and is available immediately, so the first waves and the tests run over IPSec tunnels with High Availability; several tunnels run in parallel for parallel waves. The bulk waves use Cloud Connect, a dedicated line to the Frankfurt data center that is ordered months ahead (Unit 3.6). Two routing rules follow from the static routing model: every wave's route changes are scripted and reversed in the rollback, and the on-premises and cloud subnets must not overlap.
The DNS side is prepared in advance. Lower the TTL of every record that a wave moves to 60 seconds at least one old TTL before the cutover, so that clients follow the change within a minute, and raise it again after the wave is stable (Unit 3.7).
4. The Database Wave
Databases that replatform to DBaaS move by dump and restore, which makes them a wave of their own with a hard write freeze. The sequence for FinCorp's PostgreSQL databases is:
- Prepare the target. Create the DBaaS for PostgreSQL cluster with its replication mode, pooler and off-site backup location (Unit 5.3). Create the users and groups on the target first: a standard dump includes GRANT statements but not user definitions, and superuser access is not available, so
pg_dumpalldoes not work. - Freeze writes. Stop the application or set it read-only. The outage starts here.
- Dump and restore with
pg_dumpandpsqlorpg_restore, from a host that reaches both sides over the hybrid link. - Verify row counts and application checks against the frozen source.
- Switch the application to the cluster's DNS name. The outage ends here.
MongoDB follows the same shape with mongodump and mongorestore, with one stricter rule: --oplog cannot be used because it requires elevated privileges on restore, so no write may happen during the dump, or the dump is inconsistent. The downtime of a database wave is the dump plus the restore plus the verification, which depends on data size and on the link: measure it on a full-size rehearsal copy, never estimate it from a sample.
Databases that run inside VMs on the VMware path, such as the core-banking systems, move with their VMs through VCDA and keep their own engines. Only the replatformed databases take the dump-and-restore wave.
5. Validation Gates and Rollback
5.1 Gates
Every wave passes the same gates, and a failed gate triggers the rollback rather than a fix under pressure:
| Gate | Before cutover | After cutover |
|---|---|---|
| Data | Rehearsal restore timed; source backup taken | Row counts or checksums match the frozen source |
| Function | Smoke tests written | Smoke tests pass on the target |
| Operations | Monitoring pipelines and dashboards ready (Unit 7.2) | Metrics and logs arrive; alerts are green for the agreed window |
| Network | Routes and DNS changes scripted, TTL lowered | Dependency calls succeed within their latency budget |
| Sign-off | Go decision by the application owner | Acceptance, after which the source may be retired |
5.2 Rollback Per Path
The cheapest rollback is a source left untouched. Keep every source VM stopped but intact until the target passes its post-cutover gates, and keep its database read-only rather than deleted. Rolling back is then a reversal of the route and DNS changes and a restart of the source.
On the VMware path, VCDA reverses failovers, so a migrated VM can be returned to the on-premises site.
On VDC targets, snapshots are the rollback point for changes after the first boot, such as driver installations or configuration fixes. Create a snapshot of the volume before each risky change. A snapshot copies the whole volume capacity, is not incremental, stays in the data center location where it was created, and is never deleted automatically. Data still in the VM's RAM is included only if you flush it with sync or shut down the guest first, and a snapshot of a running database is not transaction-consistent. To roll back, provision a new volume from the snapshot, set it as the boot device, and delete the snapshots once the wave is accepted, because they keep costing until someone does.
One rollback is not free: once the target has accepted writes, returning to the source loses them unless they are copied back. Every wave therefore has a point of no return, written into the runbook, after which the plan rolls forward.
6. FinCorp's Wave Plan
FinCorp combines the dispositions, paths and gates into one plan:
| Wave | Scope | Path | Link | Downtime per service | Rollback |
|---|---|---|---|---|---|
| 0 | VCDA pairing, Backup Units, monitoring, TTLs lowered, address plan checked | None | VPN tunnels | None | Not needed |
| 1 | Internal tools, about 40 VMs | Image conversion to fc-prod-fra-core |
VPN tunnels | One maintenance evening each | Restart source |
| 2 | Payment switch, Windows and BIOS-based | Backup and restore onto blank Dedicated Core servers on fc-data |
VPN tunnels | Freeze to first boot, in one weekend window | Restart source |
| 3 | Portal database | Dump and restore into fc-pg-portal |
VPN tunnels | Freeze, dump, restore, verify; rehearsed | Source read-only, re-point the portal |
| 4 to 9 | Core-banking dependency clusters, one per wave | VCDA into the Private Cloud in Berlin | Cloud Connect, VPN as backup | Final synchronisation and boot | Reverse failover |
| 10 | Source decommissioning | None | Links reduced | None | Not possible: point of no return |
The order follows the risk. Wave 1 tests the runbook on systems that can tolerate a long evening. Wave 2 exercises the hard guest changes on a contained set. Wave 3 moves the first regulated data under a measured write freeze. The core-banking waves come last, on the path that changes the least in the guest and with the shortest windows. After every wave, FinCorp's continuity plan from Unit 7.1 and the observability planes from Unit 7.2 are extended to the moved systems before the next wave starts.
Decision Summary
| Question | Choose | Hard constraints |
|---|---|---|
| VM must keep its VMware operating model | VCDA into Private Cloud | Pairing and firewall paths (443 outbound on-premises, 55443 at the cloud site); protection billed per VM |
| Small or drainable VM onto VDC compute | Image conversion and upload | Supported disk formats only; images bound to the upload region; dynamic disks expand to full size |
| Firmware or CPU change, or a Windows estate onto VDC compute | Backup and restore with the Acronis agent | Agent needs internet and its ports; changes after the last backup are lost unless frozen |
| Database moving to DBaaS | Dump and restore with a write freeze | Users created first; no superuser, so no pg_dumpall; MongoDB dumps without --oplog |
| Grouping and ordering | One dependency cluster per wave, least critical first | Do not split a cluster across the hybrid link |
| Rollback | Untouched source, reverse failover, or snapshot before each change | Snapshots are location-bound and not transaction-consistent |
Summary
A large VMware migration onto IONOS CLOUD is engineered per workload. Each workload first gets a disposition, then a path: image conversion for small, drainable VMs onto VDC compute, VCDA replication into Private Cloud for VMware-bound systems with short windows, and Acronis backup and restore for firmware and driver changes. Databases replatform by dump and restore under a write freeze. Waves follow dependency clusters, ride the link their volume needs, and pass gates with a rollback that is cheapest when the source stays untouched until acceptance.
Key Points:
- Decide rehost, replatform or refactor before choosing a path; the disposition sets the risk.
- Image conversion takes full-size disks per region and needs VirtIO and a UEFI fallback boot path; downtime is the whole copy.
- VCDA replicates asynchronously, migrates free of charge and reverses failovers; downtime is the final synchronisation.
- Backup and restore adapts boot partitions and VirtIO drivers; freeze the source before the last backup.
- Every wave has gates, a rollback, and a written point of no return.
Important Terminology:
- Disposition: the decision whether a workload is rehosted, replatformed, refactored or retired.
- Dependency cluster: a set of workloads that must move together because they call each other within a tight latency budget.
- Point of no return: the moment after which the target holds writes the source does not, so the plan rolls forward instead of back.