An IT migration is the transfer of an element of the information system (an infrastructure, an application or data) from a source environment to a target environment, until the source is shut down. The word therefore covers three objects, which involve neither the same risk nor the same teams.
The decision follows an order. First name what is really moving: the platform, the software, the data, or several at once. Then choose how to cut over and how to roll back. Finally, check that the timing is right. A migration with neither an imposed deadline nor a named benefit can wait, and one whose source will never be switched off leaves two systems to maintain.
Infrastructure migration, application migration, data migration
| Type of migration | What changes | What must remain identical | Dominant risk | Evidence expected before the cutover |
|---|---|---|---|---|
| Infrastructure | Servers, storage, network, place or mode of hosting | The software and the data | Degraded performance, forgotten network flows, licences tied to hardware | Load and disaster recovery tests on the target |
| Application | The software: major version, new software package, rewrite | The service delivered to the business | Functional gaps, broken interfaces, disrupted working habits | Business acceptance testing and comparison of results with the old system |
| Data | The medium, the format or the model of the data | The meaning and completeness of the data | Losses, duplicates, erroneous conversions | Reconciliation between source and target, validated by the business |
A real project often combines several rows of the table. Replacing a management software package changes the application and requires migrating the data. Changing hosting provider while changing database engine combines infrastructure and data.
The rule of prudence is to count the types moved at the same time. When two of them change on the same day, a gap observed after the cutover becomes hard to attribute: the cause may lie in the new platform, the new software or the conversion. If the schedule allows, it is better to separate them: move as is, stabilise, then transform.
Two cases have their own rules, covered elsewhere on this site. When the target is a cloud provider, the decision is taken application by application, as described on our page on cloud migration. When the software itself is reworked in depth, the workstream falls under application modernisation.
The criteria for deciding
The reason comes first. An imposed deadline (end of support for a component, end of a hosting contract, regulatory obligation) sets the date and argues for the narrowest possible migration. A chosen reason, such as lowering an operating cost or reducing technical debt, leaves time. In return, it requires a named benefit that can be observed afterwards.
The business’s tolerance of interruption comes next. How long can the service stop, and what volume of operations can be re-entered after an incident? These two answers determine the cutover strategy more surely than technology does.
Third criterion: knowledge of the existing system. An application whose interfaces, batch processes and user-made exports nobody can list any more cannot be migrated without a prior inventory. The forgotten dependencies are the ones that show up the day after the cutover.
Then there is the capacity of the teams. Operations must be able to run the target from day one, and users must have been prepared when their screens change. This preparation, including user training, falls under change management and is planned together with the cutover.
Single, batch or parallel cutover
A single cutover moves the entire scope on a given date, during an interruption window. It avoids any coexistence and requires only one data migration. On the other hand, it concentrates the risk on a few hours and requires a global rollback.
A batch cutover splits the scope by site, by entity or by functional domain. Each batch exposes fewer users and informs the next. The price to pay is coexistence: temporary interfaces between the old and new systems, and, for each shared data item, a written answer to the question of which of the two is authoritative.
Parallel running has the same operations processed by both systems for a bounded period, then compares their results. It is solid and costly evidence, since each operation is processed twice, often at the price of double data entry. It is reserved for processes where an error is expensive, such as payroll, invoicing or accounting.
For an infrastructure, a fourth way exists when the architecture allows it: directing a growing share of the traffic to the target and observing before widening.
Whatever the strategy, the cutover is rehearsed as a dry run. The rehearsal reveals the forgotten manual steps and shows whether the sequence fits within the interruption window.
Data migration sets the schedule
Data migration is an often underestimated workstream, and it is frequently what dictates the cutover date. It begins with a scope decision: migrate the entire history, or only the live data while archiving the rest. Everything that is not migrated must remain accessible for as long as retention obligations require.
Next come the mapping rules between the old model and the new, then cleansing. Duplicates and inconsistent values tolerated for years by the old system may be rejected by the new one. Correcting them at source, before the migration, costs less than handling the rejects during the cutover window. This work draws on the practices described on our page on data quality.
Each data migration rehearsal is measured: duration of extraction, conversion and loading on real volumes, number of rejects, differences in counts and totals between source and target. The measured duration sizes the interruption window. The reconciliations, for their part, are signed off by the business, the only party able to say that a migrated record is correct.
Test data sets drawn from production often contain personal data. Their use falls under the GDPR, Regulation (EU) 2016/679, during the migration as in ordinary times.
Rollback and its point of no return
Rolling back remains simple as long as the source is intact and the target has received nothing that the source does not know about. As soon as users enter data into the new system, rolling back means transferring these entries to the old one or losing them. This moment is the practical point of no return of the cutover. It generally falls at the reopening of the service.
The rollback plan therefore answers three questions, written in advance. Until what time can you roll back without loss? Which observations trigger the rollback: duration exceeded, data migration check failed, critical function unavailable? Who takes the decision, with what delegation in the middle of the night?
Beyond this point, unless the entries are transferred to the old system, what remains is to correct on the target or to run in degraded mode while correcting. The source is then kept read-only for a period set in the plan, then switched off.
When to refrain from or postpone an IT migration
Refraining is the right decision when the element to be migrated is due to be retired or replaced shortly, or when nobody can say who will benefit from the result. A purely technical migration, with no beneficiary, consumes the budget and attention that the next one will need.
Postponing is justified in specific situations: a period of high business activity (closing, seasonal peak), an unfinished inventory of dependencies, the impossibility of rehearsing the cutover, the unavailability of the people who know the existing system, or another large-scale change imposed on the same users. A serious postponement names the condition that will lift it and the review date.
Postponement also has a cost when the reason is imposed. A component out of support keeps running without patches, and extending a contract that has reached its term is negotiated in a hurry. The decision therefore compares two risks, that of migrating now and that of waiting.
For each migration under consideration, the decision file brings together what is moving, the way of cutting over, the deadline for rollback and the shutdown date of the source. As long as a project cannot answer on these points, its place is in study and its entry in the schedule can wait.