Skip to content

Migration · Cloud architecture · 7 min read ·

Migration by disposition, not by lift

A cloud migration that treats every workload the same ends up moving the problems with the servers. How to classify an estate into dispositions, plan the waves, and keep the business running while the ground moves.

"Lift and shift" is a disposition, not a strategy. It is the right answer for some workloads and the wrong one for most, and a migration plan that applies it to everything inherits every operational problem of the old estate at a higher monthly price. The alternative is not a grand rewrite; it is an honest classification of what you have, and a plan that treats each class differently.

Start with an inventory you can trust

Discovery tools produce a list of servers. A migration needs a list of workloads: the application, its data, its dependencies, who owns it, who uses it, and what happens to the business if it is down for an hour. Build that list by talking to owners, not by scanning ports, and expect it to be shorter than the server count and longer than anyone's mental model. The workloads nobody can name an owner for are the first candidates for retirement.

For each workload, record three facts that drive the disposition: how much it changes (frozen, maintained, actively developed), what it depends on (a licensed database, a specific OS, hardware, another workload), and what it costs today, including the people.

The dispositions

Retire. Nothing to migrate. Workloads with no owner, no users or a replacement already in place. Every estate has some; finding them before the migration is a saving with no cloud bill.

Retain. Stays where it is, for now. A mainframe, a regulatory constraint, a workload scheduled for replacement next year. Retaining is a decision with a review date, not a default.

Repurchase. Replaced by a managed service or a product: the self-hosted mail server, the home-grown ticketing system, the database whose only role is to feed a report. Often the largest cost reduction in the plan, and the one that needs the most change management.

Rehost. Moved as-is. Right for frozen workloads with no cloud dependency problems, where the goal is to vacate a data centre on a date. Plan to replatform them afterwards or accept they will cost more than they did.

Replatform. Moved with targeted changes: the database to a managed service, the file share to object storage, the VM to a container image, the cron job to a scheduled function. Most of the estate lands here. The changes are chosen because they remove operational burden, not because they are interesting.

Refactor. Rebuilt for the platform. Reserved for workloads that are actively developed, strategically important, and constrained by their current shape. Expensive, slow, and worth it for a handful of systems; disastrous as a default.

Plan in waves, by dependency

Workloads move in groups that make sense together: a system and the things that talk to it, so that no wave leaves a chatty connection crossing the network between the old estate and the new one for months. The first wave is the one that proves the mechanism with the least risk: a non-critical workload with a real owner, enough dependencies to exercise the networking, and a clear rollback. The last waves are the ones that need the most change, by which time the landing zone, the pipeline and the team's habits are proven.

Every wave has a cut-over plan, a rollback plan that has been rehearsed, and a success measure that is about the business (orders processed, reports delivered) rather than about servers being green.

Data is the schedule

Compute moves in minutes; data moves at the speed of the link and the tolerance for downtime. For every stateful workload decide early between a one-time copy with a maintenance window, continuous replication with a short cut-over, or dual-running with reconciliation. The choice sets the wave's duration and often the disposition: a database that cannot tolerate a window is one to replatform onto a managed service with built-in replication rather than to rehost.

What breaks that nobody planned for

  • Hard-coded addresses and hostnames, in configuration, in scripts, in other systems' allow-lists. Inventory them before the wave, not during the cut-over.
  • Licensing that is tied to cores, sockets or a specific hypervisor, and costs more in the cloud than the compute does.
  • Latency between a moved workload and one that stayed, which was invisible on a LAN.
  • Backups and monitoring that covered the old estate and nobody re-pointed.
  • Identity: service accounts, certificates and shared credentials that were provisioned by hand ten years ago and documented nowhere.

Measure the thing you were promised

A migration is sold on cost, agility or risk, and it is held to account on whichever one was promised. Decide the measure before the first wave, baseline it, and report it after each wave. If the promise was cost, the saving usually arrives through retirement, repurchase and rightsizing after the move, not from the move itself; say so up front.

A cloud migration engagement starts with the inventory and the dispositions, because the plan that follows is only as good as the classification under it. The landing zone is the thing to have in place before the first wave lands.