Every team has backups. Far fewer can say when they last restored one, and almost nobody can say how long it took.

That gap is where the real risk lives. A backup job that reports success is telling you a job finished, not that your data is recoverable. The two are different claims, and only one of them matters at three in the morning.

What follows is a restore drill small enough to run in an afternoon, with no new tooling and no project plan. It is written for a team that runs a SaaS product on a cloud provider and has never done this deliberately.

Start by writing down what you think will happen

Before you touch anything, write three numbers down somewhere your team can see them later.

How long you think a full restore of your primary database will take. How much data you expect to lose, measured in minutes of writes. Which person you think would run it.

Those three guesses are the point of the exercise. The value of the drill is not that the restore works, it is the distance between what you guessed and what actually happened. Teams that skip this step end up with a restore that technically succeeded and no way to tell whether that was good or bad.

Pick the one system that would actually end you

Not the whole estate. One system, chosen on a single question: if this were gone for a day, would we still have a company?

For most SaaS teams that is the primary application database. Occasionally it is an object store full of customer uploads, or the identity provider, or a queue holding work nobody logged elsewhere. Pick it honestly rather than picking the one that is easiest to test.

Everything else can wait for the second drill. One system restored properly teaches you more than five systems half tested.

Restore into somewhere new, never over the top

Restore to a fresh instance with a new name, in an environment that is not production and not the environment your engineers are working in that day.

This matters for two reasons. The obvious one is that an in place restore during a drill is how a drill becomes an incident. The less obvious one is that restoring into a new target is the only way to find out whether your recovery procedure depends on things that only exist in production: a security group, a DNS name, a secret, a role, an extension someone installed by hand in 2024 and never wrote down.

Those dependencies are the usual reason a restore fails under pressure. The data comes back fine and the application cannot reach it.

Time it with a clock, not an impression

Start a timer when the decision to restore is made, not when the restore command runs. The gap between those two moments is often the largest part of the outage, and it is invisible unless you measure it.

Record four timestamps. When you decided. When the restore started. When the data was readable. When the application was serving correctly against it.

The last one is the only number that means anything to a customer, and it is reliably the one teams do not have. Plenty of organisations know their database restores in forty minutes and have no idea that pointing the application at it takes another two hours of configuration nobody had written down.

Verify the data, do not just admire it

A restored database that opens is not a verified restore. Run three checks.

Count rows in your highest volume table and compare against production. Pick five recent records you can recognise and confirm they are there and correct. Then run one query that exercises a relationship across tables, because referential damage is the failure mode that looks fine until it does not.

If you have a schema migration history, check which migration the restored copy is at. A restore that lands you three migrations behind the running application is a restore that will fail in an unhelpful way.

Write down what broke, and fix one thing

Something will break. That is a successful drill, not a failed one.

The failure is usually in one of four places: a permission the restoring account did not have, a secret or connection string that pointed somewhere that no longer exists, a manual step nobody documented, or a dependency on a resource that was never in version control.

Pick the single worst one and fix it this week. Then put the drill in the calendar again for a quarter out. Two drills a year, each fixing one real defect, beats an annual exercise that produces a document.

Make the next one cheaper than this one

The reason restore drills stop happening is that they stay expensive. Each one should leave behind something that makes the next one shorter.

Script the restore steps you did by hand, even badly. Put the four timestamps in the same file every time so you can see the trend. Keep a short list of the dependencies the drill exposed, and move them into infrastructure as code as you go.

After three rounds the drill becomes a command and a coffee, and at that point you can start testing the systems you deferred.

Where this sits in the frameworks, briefly

If someone later asks you to evidence this, you are already most of the way there. Recovery testing appears in every mainstream control set, usually asking for the same two things: that restores are tested periodically, and that the results are recorded.

The timestamps, the verification checks and the list of what broke are that record. Produced as a by product of doing the work, which is the only kind of evidence that stays true.

For a wider map of the control set this sits inside, IRM's free CIS Controls gap assessment covers CIS Controls v8.1 at IG1, IG2 or IG3 scope, up to all 153 Safeguards, and returns a risk ranked remediation roadmap. Data recovery is one control among those, and seeing it next to the others tends to settle arguments about sequencing.

About IRM's cloud security work

Disclosure first, so you can weigh the advice accordingly: this post comes from IRM Consulting & Advisory, and the links in it point to our own pages.

The pattern behind this post is one we see repeatedly. Backups are configured competently, monitoring on the backup job is green, and nobody has ever measured the path from decision to serving traffic. The first time that path gets measured is during the incident, which is the most expensive possible moment to discover that it runs through a person on holiday.

IRM Consulting & Advisory provides cloud security consulting and cloud security management to SaaS businesses in Canada and the United States.

One question for anyone who has done this more than twice: what finally made the drill stick for your team, a calendar invite, a customer asking for evidence, or an outage that made the argument for you?