The instinct when planning disaster recovery is to protect everything. It feels responsible, it avoids an argument about whose system matters less, and it is the most reliable way we know of ending up with no disaster recovery at all.
The reason is arithmetic. Replica storage, restore-point overhead and replication licensing all scale with what you include, so an estate that protects all forty machines equally produces a quote that nobody signs. The project is shelved, and it gets done a year later under worse conditions — usually just after something has already gone wrong.
Start From Services, Not Virtual Machines
Open the hypervisor console and you see virtual machines. The business does not have virtual machines; it has things it does. Invoicing. Taking orders. Paying people. Answering customers.
So the unit of classification is the service, and the machines follow from it. This matters more than it sounds, because of the dependency trap: a tier-one application sitting on a tier-three database recovers into nothing. The application comes up, finds no data, and you have paid for a replica that cannot work. Classify the service, then pull in every machine it needs — the app, the database, the queue, the thing that authenticates it — and give them all the same tier.
Write the dependencies down while you do it. Most estates discover at least one surprise, and an incident is an expensive place to find it.
A Worked Classification: 40 Virtual Machines
Here is a realistic mid-sized estate, sorted. The counts are illustrative; the shape is not.
| What it is | VMs | Tier | Why |
|---|---|---|---|
| Identity, DNS, DHCP | 3 | Critical | Nothing else authenticates or resolves without them. Recovered first, always. |
| Finance / ERP and its database | 4 | Critical | Revenue and payroll stop. The business notices within an hour. |
| Customer-facing web and API | 4 | Important | Outside parties see it, but a few hours of a status page is survivable. |
| Line-of-business applications | 5 | Important | Staff cannot do their jobs, but the work queues rather than disappears. |
| File services and email gateway | 3 | Important | Painful and visible; not immediately expensive. |
| Internal tools, wiki, ticketing, monitoring | 11 | Standard | Rebuild from backup over a day or two. Nobody loses money waiting. |
| Dev, staging, build agents, migration leftovers | 10 | Leave | Recreate from configuration when there is time. Protecting these is where budgets go to die. |
Seven machines replicate continuously. Twelve more replicate but can boot cold. Eleven are a backup-driven rebuild. Ten are not protected at all. The DR bill is now a fraction of the “protect everything” version, and — this is the part people miss — the test is now small enough that you will actually run it.
What People Wrongly Include
- Dev and staging. The most common line item in an over-sized DR design. If production is down, nobody is developing against staging anyway. Recreate it afterwards.
- Build agents and CI runners. Stateless by design. Rebuild them from the definition that already exists in your repository.
- Monitoring that watches the dead site. Worth thinking about rather than replicating reflexively — monitoring that lives inside the failed environment tells you nothing during the incident. What you want is monitoring from outside, which is a different purchase.
- Second copies of SaaS. If a service is someone else’s to run, your DR plan for it is their SLA plus an export, not a replica.
- Machines nobody can name. Every estate has them. An unidentified VM is not a reason to replicate it; it is a reason to find out what it does before the next review.
What People Wrongly Leave Out
This is the more expensive mistake, because it is only discovered during the recovery. The pattern is always the same: infrastructure that is boring, that nobody owns, and that everything else quietly depends on.
- Domain controllers and identity. The single most common omission. The applications come up and nobody can log in, including you.
- DNS and DHCP. Recovered systems that cannot resolve each other are not recovered.
- Licence servers. Some applications will not start without one, and a licence tied to hardware that no longer exists is an incident inside an incident.
- The jump host or management VM. If the only route into the environment lived in the environment, you now have a working estate you cannot reach.
- The file share holding the runbooks. Documentation stored inside the thing that is down. Worth checking today rather than on the day.
- Certificate authorities and secrets managers. Internal PKI and vaults are infrastructure, not applications, and they are rarely on anyone’s list.
- The backup server itself. If the recovery plan for the standard tier is “restore from backup”, then the backup infrastructure is a critical system, not a standard one.
What “Leave It” Actually Means
Excluding a machine from disaster recovery is not the same as abandoning it. Everything in the bottom two rows should still be backed up, and the rebuild route should be written down — which image, which configuration, roughly how long. “We would rebuild it” is a plan only if someone has checked that they could.
It is also a decision the business should see, not an engineering convenience. The list of what is deliberately not protected belongs in front of whoever owns the consequence, once, in writing. That conversation is uncomfortable for ten minutes and saves a much worse one later.
Review It When the Estate Changes
Classifications rot. A system moves from internal tool to business-critical over eighteen months and nobody re-tiers it; a service gets decommissioned and keeps being replicated for two years. Review when something material changes — a new product line, a new regulator, an acquisition — rather than annually out of habit.
Doing This With Us
The classification is the first deliverable of an assessment, and it is the part that saves the most money, because it is where “forty machines” becomes “seven, twelve and the rest”. The DR planner will show you what your estate implies before you talk to anyone: replica footprint, seed time over your own uplink, the bandwidth replication needs afterwards, and what your recovery target forces to be running already.
Next in the series: where the second copy lives — the 3-2-1 rule rewritten for the UAE, and the cross-border question that comes with it.
If you would rather talk it through, our DRaaS service page sets out the shape of the engagement, and telling us what you run gets you an assessment call.


