Skip to content
My accountDeutsch

PracticeReading time 6 min.

RPO and RTO: set recovery objectives before something fails

How much data can an outage cost, and how long can it last? If you only answer both questions in an emergency, you will find that the technology has given a different answer.

Managed by the Operations department

Three storage media side by side on a light background: a grey tape cartridge, an SSD in an aluminium case and a hard drive with a visible circuit board.

At a glance

  • The RPO limits data loss, the RTO the downtime โ€“ both are derived per application from the consequences of an outage.
  • Backup, snapshot and replication protect against different types of damage; replication also transfers the wrong write operation.
  • A recovery objective is only valid once a rehearsal has achieved it under the stopwatch.

Two numbers, two questions

The Recovery Point Objective (RPO) is the maximum data loss you accept, measured in time. An RPO of one hour means: after a recovery, at most the changes from the last hour before the outage are missing.

The RTO (Recovery Time Objective) is the downtime you set as a goal; it is below the maximum tolerable downtime, after which damage can no longer be recovered. It begins with the outage and ends when the application is working correctly again from a business perspective โ€“ not when the server has started.

Data loss = time of outage โˆ’ time of last usable data state โ‰ค RPO

Downtime = first functionally correct response โˆ’ time of outage โ‰ค RTO

The two numbers are independent of each other. A nightly backup might be restored in an hour and still lose a whole working day. A mirrored database restarts in seconds and has still lost an accidentally deleted record. Each number answers its own question and requires its own procedure.

Derive objectives from the consequences โ€“ per application

A blanket objective for all systems is either too expensive or too weak. An ERP system, a customer portal and an archive of old projects have different consequences if they fail. The usual procedure for this is called a business impact analysis; ISO 22301 requires it for a business continuity management system, and BSI Standard 200-4 describes it in detail.

For each application, you answer four questions:

  • What is lost if an hour of data is missing, and can it be re-entered? Orders from emails can be entered again, a payment or a series of measurements often cannot.
  • At what duration does damage occur that can no longer be recovered โ€“ a missed deadline, a closed shop, a contractual penalty?
  • What obligations apply? Art. 32 GDPR lists among the protective measures the ability to quickly restore the availability and access to personal data after an incident, and a process for regularly testing their effectiveness.
  • What does the application depend on? A portal that does not start without a login service cannot achieve a shorter RTO than this service.

The answers result in a pair of RPO and RTO per application. Only then do you choose the procedure โ€“ not the other way around. If you buy the technology first, you get the objectives it happens to come with.

Restoring the data is only one part of the recovery time; detecting, deciding and testing also cost time. The example works with set assumptions, not with measured values.

Example: recovery with set assumptions
Detect and isolate outage
20 minutes
Decide on and trigger recovery
10 minutes
Provision target system
30 minutes
Restore 540 GB at 150 MB/s
60 minutes
Functionally test and approve
30 minutes
Total recovery time
150 minutes (2.5 hours)

Restoring accounts for only 40% of the time in this example. If the application requires an RTO of two hours, this plan misses it by 30 minutes, even though the restore itself only takes 60 minutes: the data is only back 120 minutes after the outage, and the testing comes after that.

Backup, snapshot and replication: what each achieves

All three procedures create a copy, but protect against different types of damage; none replaces the other two.

Three procedures and what they protect against
MethodProtects againstDoes not protect againstPrimarily determines
BackupDeletion, corruption, loss of the entire systemData loss since the last backupRPO via the interval between backups
SnapshotFailed updates, user errorsFailure of the storage it resides onShort RTO for rollbacks
ReplicationFailure of a node or locationIncorrect write operations โ€“ they are also transferredLow RPO and short RTO

A backup is a copy that is kept separate from the original and retained for a period. Its RPO is the interval between two backups, and its recovery takes longer the more data is restored.

A snapshot captures the state of a volume or datastore and usually resides on the same storage system. It is created and rolled back in seconds, but shares any damage to this system. A snapshot of a running database is also only crash-consistent if the application was not previously put into a consistent state.

Replication transfers every change to a second system. It keeps the RPO low, but transfers every write operation, even incorrect ones; the Private Cloud recovery page therefore explicitly states that none of its tiers replaces a backup.

A journal combines both advantages: a base backup and the continuous log of all transactions allow returning to any point in time still covered by the journal.

With ENTRONYX managed databases, for engines with a primary node and two or more nodes, a replica takes over the write role if the primary node fails, typically in under 30 seconds; your application must be able to reconnect for this.

three of the nine engines recover to the second, because they keep a continuous journal:

  • PostgreSQL
  • MySQL
  • MongoDB

The others restore to their last backup, snapshot or export โ€“ except for Apache Kafka, which retains events in the topic instead of recovering them. How far back the journal reaches per engine is shown in the table in the Daily full, journal in between section.

Retention: as long as an error can go unnoticed

The RPO tells you how fresh the latest state is. Retention tells you how far back you can go โ€“ and it must be longer than the time an error goes unnoticed. Corruption that is only noticed after three weeks can no longer be fixed from a fourteen-day retention.

For managed databases, retention ranges from 7 to 365 days depending on the selected tier; the tier from which the backup is additionally stored in a second region is stated on the page itself.

A backup that an attacker can delete with the same credentials does not protect against ransomware. The Backup Vault stores backups georedundantly and immutably: during a lock period of up to 10 years, it rejects the deletion and overwriting of a locked version with status code 403.

If you have to retain data for a long time, you also pay attention to the way back. A retrieval from the Cold Archive takes three to five hours until the first bytes. This is suitable for retention obligations, but not as the only copy of an application with a short RTO.

Rule of thumb

Three copies of your data on two different storage types, one of them in a different location โ€“ known as the 3-2-1 rule. At least one of these copies should additionally be immutable.

Test recovery, do not assume

A recovery that has never been tested is an assumption. Only the test shows whether the runbook, access and schedule hold up and whether something is missing that lies outside the backed-up environment.

The Private Cloud disaster recovery test page describes a procedure that can be applied to any environment. The startup takes place in an isolated network with no path to production, and from the state that is already in the destination anyway โ€“ without a specially prepared snapshot. The clock runs from the trigger to the first functionally correct response of the application; detection and decision-making precede this and are added to a guaranteed time. You define the functional checkpoints yourself beforehand.

A test is only passed there if all three criteria are met: time kept, data state kept, functionally passed. A failed test is repeated within 30 calendar days, and until the successful repetition, the time commitment of the tier is considered suspended. On the same page, the disaster calculator simulates a failure, and the destination region section lists which destinations the tiers currently have.

The same prerequisites apply to your own test: a runbook that works without its author, named dependencies such as directory service and name resolution, and credentials that are not only located in the environment to be recovered.

Checklist

Go through these steps for each application individually.

  1. List all applications with their dependencies.
  2. Define RPO and RTO per application based on the consequences of a failure.
  3. Choose the procedure per destination: backup, snapshot, replication or journal.
  4. Set the retention longer than the time an error can go unnoticed.
  5. Keep at least one copy immutable and in a different location.
  6. Calculate the recovery time including detection, decision-making and testing.
  7. Test the recovery with a stopwatch and record deviations in writing.
  8. Review the targets again as soon as an application or its data volume changes significantly.

Published on 8 November 2026. Prices and product values in this post are taken from the price list and reflect their current status.

Further reading

Read more

Questions for the department

Technical questions about a post are answered by the person responsible for the department โ€” in a conversation, not a presentation. For quotes and framework agreements, sales will lead the initial conversation.