1. The hidden risk of concentration
Many companies keep all their systems with a single provider, in a single region. When that works it is convenient, but when an area, a central service or a contract fails, everything stops together. The risk is not only technical: if you have no alternatives, a price increase or a unilateral change to the terms is hard to manage, because the cost of changing course grows with time.
Distinguishing resilience from sovereignty is the first step. Where data is stored and who can legally access it are different questions. A provider with data centres in Europe may be subject to legal obligations from other countries, and vice versa. Resilience is built through redundancy and tested recovery procedures, not through a geographic label on the contract. Before choosing, ask the provider where the data is, how it is replicated and what the procedures are if an entire area fails.
2. Hybrid architectures: what to distribute and where
A hybrid model combines on-premise resources, meaning on the company's premises or in a dedicated colocation facility, with one or more cloud services. The basic criterion is to separate workloads by criticality, data sensitivity and latency requirements. Systems that must respond to production in very short times stay close to users, while historical archives and backup copies can be held elsewhere. Each choice should be justified by a concrete requirement, not by an abstract preference for a technology.
Not everything should be distributed. Every additional component adds complexity, network costs and skills to manage. An SME often gains more resilience from two well-tested copies of a system than from four poorly documented environments, because real resilience depends on the ability to recover, not on the number of environments. Sizing should be based on critical processes, not on enthusiasm for the architecture, and every environment added must have a clear owner.
3. Four models compared
The table compares the most common models. None is right in absolute terms: the choice depends on the maximum acceptable downtime for each service, the sensitivity of the data and the budget you can devote to management. A company with few critical systems can cover most of the risk with the second model, while a company with processes that must always run needs a more articulated approach.
The model with two independent providers reduces the risk of systemic blackout the most, but it is also the most demanding to maintain, because procedures, monitoring and testing must be carried out on two different platforms. For many SMEs the most sensible starting point is a combination of on-premise systems for the most sensitive data and a backup copy with a second provider, with a tested recovery plan.
| Model | Advantages | Limits | When it makes sense |
|---|---|---|---|
| Single provider, single region | Simple, with contained management costs | A regional failure stops everything | Low-criticality workloads with external backup |
| Single provider, multiple regions | Geographic redundancy without changing provider | Contractual and technical dependence on the same provider | Applications that must stay in service |
| Two independent providers | Reduces the risk of systemic blackout | More complex management and data to keep aligned | Critical systems and last-resort copies |
| On-premise plus cloud hybrid | Direct control over the most sensitive data | Requires in-house skills or a partner | Regulated data and low-latency workloads |
4. Disaster recovery: redundancy is not backup
Replicating a database to another site also means replicating a deletion error, if you do not have copies with history. An operator who accidentally deletes a table sees the deletion propagate in real time to all synchronised copies. You therefore need independent copies, with versions over time, and two values to agree with the business before choosing the technology: how much data you can afford to lose at most, and how long you can stay down before the damage becomes serious.
The 3-2-1 rule remains a good starting point: three copies, on two different media, one of them off site. It must, however, be complemented by periodic restore tests, run on real systems or on a test environment that mirrors production. An untested backup is a hope, not a guarantee, and testing almost always reveals forgotten details, such as expired credentials or undocumented dependencies.
5. Lock-in: how to avoid being trapped
Lock-in arises when services use proprietary features that do not exist elsewhere. Managed databases, messaging services and orchestration tools are convenient, but they can make moving a workload expensive. The point is not to avoid every proprietary service, but to know what the exit cost is and to accept it consciously, writing it into your architecture decisions. In practice, it means asking each time how much it would cost, in time and money, to move that service to another provider.
Preferring standard technologies, such as containers, compatible databases, documented APIs and open data formats, reduces the cost of exit. Ask your provider how you export data, in what format, in how much time and whether export is included in the contract. Having these answers before signing is worth more than any slogan about portability, and it is worth verifying them with a real export test on a sample dataset.
6. Getting started: a map before the architecture
The first step is not choosing two providers. It is mapping your systems and their acceptable downtimes. With that map you can decide where a second copy is really needed and where a well-run backup is enough. The map should be shared with those who run the business, because process owners know how much an outage really costs.
A good starting point is a technical check limited to a single system. If the restore test works and the process is clear, extend the model to other workloads gradually. Moving in small steps reduces the risk of building an architecture nobody knows how to run, and each step produces documentation that also serves the next decision.
- List the services and classify them by business impact in the event of an outage.
- For critical services, define in writing the maximum recovery time and the acceptable data loss.
- Map the dependencies: which providers, which regions and which contracts are involved.
- Choose the simplest model that meets those values.
- Plan a restore test on a real system and document the result.
- Ask candidate providers how they handle replication, data export and communication during an incident.
Frequently Asked Questions
Does a multi-provider cloud cost much more?
It depends on the scope. Replicating everything across two providers roughly doubles much of the infrastructure cost, while replicating only critical systems and last-resort copies has a much smaller impact. The real cost also includes management: monitoring, updates and testing must be done on every platform. It is worth comparing the spend with the damage of a prolonged outage, estimated together with the process owners.
Does multi-provider mean the data stays in Europe?
Not necessarily. Where data is stored and the provider's jurisdiction are different questions. A provider may store data in the EU and have a parent company subject to the laws of other countries. For sensitive data, ask where it is held, who can access it, how encryption keys are protected and how requests from authorities are handled.
Is a second data centre from the same provider enough?
It is a useful step against physical failures or failures in a single area, but it does not protect against configuration errors, attacks that target the account, or contractual problems with the provider. For this reason, multi-region redundancy should be combined with independent data copies and tested recovery procedures.
How is data synchronised between two providers?
Usually through asynchronous replication for non-critical data, and with more sophisticated solutions for data that cannot lose transactions. Each choice involves a trade-off between latency and consistency: the tighter the synchronisation, the more sensitive the link between the two environments becomes to distance and network failures. There is no single solution. It should be decided on the basis of the maximum acceptable data loss for each service.
How many people are needed to run a hybrid architecture?
There is no figure that fits everyone: it depends on the number of systems, the level of automation and the support model. Outsourced management with cover across shifts can avoid hiring scarce skills in-house. What matters is that every component has a clear owner and that runbooks, meaning the operating procedures, are kept up to date and tested at least once.