Management signs off on “RPO 15 minutes, RTO 4 hours for ERP”; in the infrastructure there is still a single backup job running at 23:00. This gap between the target on paper and the reality in the server room shows up at the first serious failure. Below, starting from the approved targets, we divide systems into tiers, choose the right technology for each tier (daily backup, frequent snapshots, replication, clustering, continuous data protection), and measure the actual recovery time and compare it with the target.
The technical meaning of RPO and RTO
RPO (Recovery Point Objective) expresses in time the maximum data loss acceptable after an outage. Technically it is the gap between the last valid recovery point and the moment of the incident. A backup taken once a day at 23:00 means about 24 hours of data loss for a failure at 22:59 the next day. That’s why the real RPO of a daily backup isn’t “1 day”; in the worst case it is the backup interval plus the time for the backup to complete. Do the math accordingly.
RTO (Recovery Time Objective) is the longest acceptable time for the service to become usable again. A common misconception is to equate RTO with the transfer time shown by the backup software. In fact RTO is the sum of these components:
- Detection and decision time: noticing the failure, making the recovery decision.
- Infrastructure preparation: bringing up the target host, storage, network and identity services.
- Data transfer or activation: restore, replica failover or cluster role failover.
- Application verification: services starting, database consistency, user access testing.
The target-setting process itself (business impact analysis, MTPD and approval) is a management matter and is described in the setting and approving MTPD, RTO and RPO targets article. Here the targets are assumed to be in hand, approved and in writing.
Dividing systems into tiers
If you write a separate target for every server in the server and system infrastructure, you end up with a hundred-row table nobody updates. Group systems into tiers that can be protected with the same technology. An example structure for a mid-sized environment:
| Tier | Example systems | Example RPO | Example RTO |
|---|---|---|---|
| Tier 0 | AD DS, DNS, DHCP, identity | Minutes | < 1 hour |
| Tier 1 | ERP database, production software | 15 min | 2–4 hours |
| Tier 2 | File server, email gateway, intranet | 4–24 hours | 8–24 hours |
| Tier 3 | Test, archive, reporting | 24 hours+ | Days |
The values should come from your organization’s approved targets; the table only shows the structure. The reason we keep Tier 0 separate is simple: everything else depends on it. If the domain controller and DNS aren’t up, users can’t sign in and service accounts can’t authenticate even if the ERP database opens. Without DNS, there is nothing.
Dependency order and the recovery chain
An application’s RTO can’t be shorter than the slowest component it depends on. Even if the ERP application server comes up in 30 minutes, if the database is restored in 3 hours, the application’s real RTO is at least 3 hours. Write down the dependency chain for each Tier 1 system:
- Infrastructure: hypervisor cluster, shared storage, core network and firewall rules.
- Identity and name resolution: at least one domain controller, DNS, certificate services if needed.
- Data layer: database servers, file shares.
- Application layer: application servers, license servers, integration services.
- Access layer: load balancer, reverse proxy, VPN, remote desktop gateway.
When you map out the chain, two problems come up in almost every environment. The first is a system labeled Tier 2 that is essential for a Tier 1 application to work; the typical example is the license server nobody remembers. The second is systems assumed to be restored in parallel but that in fact queue up because they share the same storage or the same proxy. Parallel on paper, serial in practice. How to document this order step by step is the subject of the preparing a disaster recovery runbook article.
Target–technology mapping table
Each technology meets a certain RPO/RTO band at reasonable cost. The table is a starting point for the decision; real values vary by product, configuration and bandwidth.
| Technology | Achievable RPO | Achievable RTO | Scenario it protects against |
|---|---|---|---|
| Daily image backup | ~24 hours | Hours (depending on data size) | Deletion, corruption, ransomware, hardware loss |
| Frequent backups / log backups | 15 min – 1 hour | Hours | Logical errors, return to a specific point in time |
| Instant recovery from backup | Backup interval | Minutes (reduced performance) | Loss of a single VM |
| Asynchronous replication (Hyper-V Replica, Veeam) | Minutes | Minutes – 1 hour | Site or host loss |
| Continuous data protection (CDP) | Seconds | Minutes | Site loss, short rollback |
| Failover cluster / SQL AG | 0 (synchronous) – seconds | Seconds – minutes | Node failure |
| Synchronous storage replication | 0 | Minutes | Storage or site loss (distance-limited) |
The real lesson of the table: high availability technologies don’t replace backup. A cluster, an AG and synchronous replication also carry an accidentally run DELETE command or blocks encrypted by ransomware to the other side within milliseconds. A replica is a faithful copy. It copies the mistake too. RPO against logical errors is provided only by backups and log backups that keep recovery points back in time. That’s why for Tier 0 and 1 systems we build two layers together: a replica or cluster for infrastructure failure, and frequent backups for logical errors. Implementation details of the replication options are in the Hyper-V Replica, Veeam and storage replication article; cluster design is described in the Hyper-V failover cluster planning article.
Balancing cost and complexity
Bringing RPO down from hours to minutes usually needs a second site or cloud target, additional licenses and constant bandwidth. Bringing RTO down from hours to minutes means idle standby compute capacity. For each tier, ask: can we meet the target with a lower-level technology, for example frequent backups and instant recovery instead of replication? If the answer is yes, don’t add complexity. Every extra layer is one more component to troubleshoot at 3 a.m.
Measuring restore time
The only way to say the RTO is met is to measure it. Transfer rates in product documentation or the duration of the backup job don’t show restore time. The restore path (reading from the repository, network, writing to target storage) hits different bottlenecks than the backup path; a deduplicated repository is fast when writing, but often not when reading randomly. A stopwatch, not an estimate. To measure:
- Fully restore a representative system from each tier on an isolated network and record the start and end times.
- Break the time into components: data transfer from the repository, VM registration and boot, services starting, application verification.
- Test concurrency: in an environment where a single VM comes back in 40 minutes, when ten VMs are restored at the same time the time may not grow linearly, but it will certainly grow.
- Veeam SureBackup or similar automated verification features can be used to regularly prove that the VM boots and its core services respond.
In a Hyper-V environment you can simply measure with PowerShell how long a VM takes to respond after a full restore:
$start = Get-Date
Start-VM -Name "ERP-DB01"
do { Start-Sleep -Seconds 5 }
until (Test-NetConnection -ComputerName "erp-db01" -Port 1433 -InformationLevel Quiet -WarningAction SilentlyContinue)
"SQL port responded: {0:N1} minutes" -f ((Get-Date) - $start).TotalMinutes
This measurement gives only the boot and service start time. The port being open doesn’t mean the application works; add the data transfer time and whether a key user can carry out a real transaction to the total RTO. How to make restore testing a regular habit is covered in detail in the 3-2-1-1-0 rule and restore testing article.
Worked example
Let’s check whether the target for an ERP database server in Tier 1 is met with the existing infrastructure. The values are assumptions; put in numbers from your own test restore.
| Item | Value | Explanation |
|---|---|---|
| Approved target | RPO 15 min, RTO 4 hours | Tier 1 |
| Total VM disk | 1.8 TB | Used blocks |
| Measured restore speed | 250 MB/s | From the test restore |
| Data transfer time | ~2 hours | 1,800,000 MB ÷ 250 MB/s ≈ 7,200 s |
| Infrastructure preparation + decision | 45 min | From the drill |
| Applying log backups | 20 min | Logs since the last differential backup |
| Application verification | 30 min | Key user test |
| Total | ~3 hours 35 min | Close to the target, little headroom |
On paper, the RTO is met. The headroom, however, is 25 minutes. If two more Tier 1 systems are to be restored at the same time, the 250 MB/s is shared and the target is exceeded. On the RPO side, a 15-minute log backup interval appears to meet the target; but if the log backups sit only on the same server’s local disk, they go with the server and the real RPO falls back to the time of the last copy. We see this mistake often in audits.
Closing the gap between target and reality
When you put the measurement results next to the target, the gap usually fits one of these patterns:
- Transfer bottleneck: The repository is on slow disks or behind a 1 GbE network. Fix: speed up the restore path (a faster repository tier, 10/25 GbE, a local copy), or bring the service up first with instant recovery and move the data in the background.
- Ordering bottleneck: Tier 0 systems come back late. Fix: an always-running domain controller at the second site, protecting identity services with replication instead of restore.
- Concurrency bottleneck: A single proxy or a single repository network connection forces all restores into a queue. Fix: size proxy and repository capacity for the disaster scenario.
- RPO gap: The backup interval is fine, but the copies are in the same place. Fix: copy log and incremental backups to a second location at short intervals.
- The target itself isn’t realistic: If the cost of meeting it technically isn’t accepted, the target must be resubmitted for approval with the measurement results. Silently counting the target as “roughly met” postpones the surprise until disaster day.
If you’d like to work together on measuring the technical reality behind the targets and reflecting the gaps in infrastructure design, you can find our approach on our backup and disaster recovery and infrastructure consulting pages. For the conceptual framework, NIST’s SP 800-34 Rev. 1 contingency planning guide for information systems is a good primary source.
Frequently asked questions
What is the difference between RPO and RTO?
RPO is the maximum acceptable data loss and depends on the frequency of recovery points. RTO is the longest acceptable time for the service to become usable again; it depends on the speed of restore, failover and verification.
If there is replication, do we still need backups?
Yes. Replication and clustering provide a fast return after hardware or site loss, but they also carry deletion, corruption and encryption to the other side. A recovery point back in time exists only in a backup.
How is RTO measured?
Actually restore a representative system from each tier in an isolated environment; record the detection, infrastructure preparation, data transfer and application verification times separately. Also test the scenario where several systems are restored at the same time.
Why can the real RPO of a daily backup be longer than 24 hours?
If the incident happens right before the next backup, the loss equals the backup interval; add to this the time for the backup to complete and the time for the copy to reach the second location. If the backup sits in the same place affected by the failure, the RPO goes back even further.