# Recovery point objective

## What it is

A recovery point objective, usually shortened to RPO, is the limit on how
far back in time a service can go when restoring data after a disruption.
An objective of fifteen minutes means that the organisation can tolerate
losing at most the changes from the final fifteen minutes before the event
under the agreed failure scenario. If the last verified recoverable state
is fifty minutes old, the result misses that objective even if the service
starts immediately.

Specify the data and the scenario alongside the number. Losing a recent
local cache may matter little when it can be rebuilt from a trustworthy
source. Losing accepted orders matters differently, particularly when a
customer has already received a confirmation. A deleted row, a full
regional outage and a compromise of operating credentials can leave
different backups available. A target attached to one data store and one
well stated scenario can be tested.

The recovery time objective measures how long the system can be unavailable
while it is restored. Recovery point and recovery time therefore lead to
different questions. One asks which state survived. The other asks when
useful service returned. A daily backup that restores in minutes may meet
the time target and still lose a day's work. A copy that retains recent
transactions may meet the point target and still take too long to verify
and serve.

## Why it matters

A backup schedule is an input to a design. The objective is a requirement for
what can be recovered. A job advertised as running every five minutes might
lag behind the application, fail without a useful alert or hand back a copy
that cannot be replayed. A team can test the difference by placing known
changes at recorded times and then tracing the last contiguous state it can
restore. Measuring the end result reveals problems which an ordinary
successful backup log cannot show.

The target should reflect what the organisation is willing and able to
reconstruct. Ask what happens to a receipt, an accepted booking or a supplier
event produced during the lost window. Some records can be replayed from
a durable upstream log. Others exist only in the application that recorded
them. Estimate the cost and risk of reconstructing each record, and give
the work an owner. A small objective on a small set of important state can
be more useful than the same ambitious number across every cache and report.

Frequent copying is only one approach. Depending on the data, a recovery
strategy may combine periodic snapshots with a transaction log, a replica,
an immutable copy or independent upstream event retention. These options
have different properties under different failures. A replica can transfer
accidental deletions as quickly as legitimate work. A separate copy helps
only if the restoration team can access, interpret and validate it when it
is needed.

## What it looks like in practice

Engineer ApS describes [backups that have actually been restored](/portfolio/backups-that-have-actually-been-restored/)
and the availability and handover terms in its [service levels](/service-levels/).
The portfolio account gives a reader evidence about exercising restoration.
The numerical example that follows is a hypothetical way to test a point
objective without assuming a target for the published case.

Assume a shop adopts a fifteen-minute RPO for accepted orders in the event
of a failed database. It snapshots the database and saves a sequence of
transaction records separately. At 14:00 a rehearsal removes access to the
primary system. Write known test orders before and after 13:45, as well as
a cancellation and a payment update. Start the recovery with the permissions
the duty team will actually have. Restore to a separate environment and
choose the latest consistent state whose data can be verified.

If that state includes all expected transactions through 13:50, the observed
gap is ten minutes and the point target is met for the rehearsal. A visible
order from 13:58 does not alone establish a two-minute gap if earlier records
are missing. Compare row identities and totals and check the dependencies an
order needs to be actionable. Preserve the restoration evidence and account
for what would happen to the orders in the remaining window. Record the time
when useful service could return separately.

A second rehearsal can change the circumstance. Remove the ordinary
operator's access to the shared control plane or simulate an accidentally
deleted item by choosing a clean point before that edit. Inspect the copies
available under that condition. Results for one scenario are useful evidence
for that scenario, and a reason to test additional ones whose dependencies
differ.

## Where it goes wrong

Teams often assume that replica lag is an end-to-end measure of loss.
An application may write order details to one store and attach evidence to
another. If their time boundaries differ, an apparently recent database
may reference files that did not survive. Include linked records and check
consistency across those boundaries in the actual recovery exercise.

A recovery run may still rely on unavailable credentials, an undocumented
encryption key or a human who remembers how to find the right copy. Give the
authorised duty role the actual runbook and exercise access with it. Write
down which timing and capabilities are assumptions. Time spent locating
copies is also useful evidence for the separate time objective.

Finally, review the target when data becomes harder to reconstruct. The
first order application may have started as a copy of a manual log, for
which a day's loss was manageable. After the system becomes the only accepted
record, the same target has a different meaning. Revisit the loss limit and
repeat the exercise with the affected process owner before treating the
old measurement as evidence for the new service.

## Further reading

- [NIST contingency planning guide for federal information systems](https://csrc.nist.gov/pubs/sp/800/34/r1/upd1/final)
- [AWS Well-Architected recovery objectives and disaster recovery](https://docs.aws.amazon.com/wellarchitected/latest/reliability-pillar/plan-for-disaster-recovery-dr.html)
- [NIST recovery-point objective terminology](https://csrc.nist.gov/glossary/term/recovery_point_objective)

<https://engineer.company/glossary/recovery-point-objective/>
