Why a cloud migration disappoints more often than it fails
A cloud migration most often succeeds on the technical level: the applications end up running with the provider. The disappointment comes from elsewhere. The recurring cost exceeds the initial estimate. The old data centre stays open, because a few applications could not leave. Teams operate the cloud as they operated their servers, with no gain in speed.
These three symptoms have a common cause: the migration was run as an infrastructure project, with a single indicator, the number of machines moved. Yet the result is decided elsewhere. It is decided in the choice made for each application, in the quality of the foundation that receives it, and in the ability to actually close what you are leaving.
Six strategies, one decision per application
The most common usage distinguishes six strategies, with variations from one source to another. None is good or bad in itself. The mistake is to apply a single one to the whole estate.
| Strategy | Principle | Choose it when | Point of vigilance |
|---|---|---|---|
| Rehosting | Move the application as it is | A deadline is imposed (end of lease, end of hardware support) and the application changes little | Its debt and its original sizing travel with the application |
| Replatforming | Replace a few components with managed services, without touching the business code | Operations are expensive and the architecture remains sound | Non-regression to be proved, increased dependency on the provider |
| Refactoring | Revisit the architecture, sometimes the code | The application is strategic and set to evolve for a long time | It is a software project in its own right, to be managed as such |
| Replacement | Adopt software as a service instead | The function is not a differentiator | Data migration, interfaces, contractual terms |
| Retirement | Shut the application down | Usage is residual or redundant | Archiving and retention obligations |
| Retention | Leave it in place | A technical or regulatory constraint requires it, or replacement is imminent | Set a review date and plan for interconnection |
Two practical remarks. Retirement is often the most cost-effective strategy and the least studied: a serious inventory regularly brings to light applications that no one uses any more. And rehosting is only acceptable if it is followed by a second phase of adjustment. Without this, you move the technical debt without reducing it. When refactoring is required, it falls under an application modernisation workstream, with its own rules.
The landing zone: what must exist before the first application
The landing zone is the ground prepared to receive the applications. It sets the account structure, identity federation, the addressing plan and interconnection with the existing estate, centralised logging, security rules applied by default and the tags that will make it possible to allocate costs.
Two excesses are to be avoided. The first is migrating without a foundation: each project then creates its own foundations, and everything has to be redone when volume grows. The second is designing an exhaustive foundation over many months before receiving a single application. We recommend aiming for a minimal but sound foundation, described as code, which is enriched at each wave according to the needs observed.
Our approach: waves ordered by risk
Our consultants work within your teams; decisions remain yours.
Grouping by affinity. Two applications that exchange constantly cannot be separated without consequences for response times and data transfer charges. Batches are built from the real flows.
Starting with what teaches the most. The first wave brings together applications of low criticality, but representative of the difficulties to come: a large database, an interface with a system that remains on premises, a batch process. It serves to run in the procedures.
Rehearsing each cutover. A successful cutover has had a dry run. Its duration is known, the rollback has been tried, and the criteria that trigger abandonment are written in advance.
Closing the source. An application is only migrated once its original environment has been shut down. The period of parallel running is bounded from the plan onwards, because it carries a double cost.
Tracking cost as you go. Each application arrives with an estimate. We compare actual cost with this estimate from the first weeks of operation, drawing on the practices described in our FinOps offering.
Reversibility is decided on the way in
Reversibility is the ability to leave a provider, for another one or for in-house hosting, under known conditions. It is eroded by each provider-specific service adopted. That is no reason to do without them: it is a reason to adopt them knowingly.
For each application, we document three levels of dependency: data (formats, volumes, exit charges and lead times), components (provider-specific services with no equivalent elsewhere) and skills. Regulation (EU) 2023/2854, known as the Data Act, contains provisions on switching between data processing service providers, to be taken into account when reading your contracts. When the sensitivity of the data dictates the choice of provider, the question falls under our page on sovereign and hybrid cloud.
What derails a migration
- Estimating costs from the existing estate. Reproducing in the cloud machines sized for an annual peak skews the calculation from the start.
- Keeping the difficult applications for last. They determine the closing date of the data centre. Their strategy must be decided at the beginning, even if their cutover comes late.
- Migrating without the operations teams. On cutover day, they are the ones who receive the alerts. Their training is part of the plan.
- Ignoring licences. Some software contracts change their terms depending on the hosting environment. The check comes before the choice of strategy.
The signs of a successful migration
We agree with you on observable indicators, tracked wave after wave:
- the share of applications whose strategy is decided and documented;
- the number of original environments actually shut down, rather than the number of applications moved;
- the gap between estimated and actual cost, per application;
- the number of incidents occurring during cutovers and the share of rollbacks;
- the share of resources deployed as code rather than by hand;
- the existence of a tested reversibility plan for critical applications.
The second indicator is the most revealing. As long as the source remains open, the migration adds costs without removing any.