Running a restore drill without the drama
Nathan Chadwick · · 8 min read · Updated
Listen
TLDRRead the short version
- Once a quarter, restore one VM, one database, and one file share to somewhere that is not production. Time it, and give every gap a named owner.
- A successful backup job is not a restore. File-level recovery is not a boot. A PaaS database is often not in the Recovery Services vault at all.
- Soft delete on current vaults is 14 days unless you extend it, up to 180. Lengthen it to a window you would actually notice.
- Lock immutability only after a restore has worked and you are sure of the retention. The lock is irreversible.
- Alert on failed jobs, on jobs that never started, and on vault replication, before an incident makes you look.

The job was green#
The backup blade is a comforting screen. Last night's job succeeded, the vault is geo-redundant, the retention says seven years, and nobody has restored anything since the person who built it left.
Azure Backup will keep taking that backup for as long as you pay it. The question a drill answers is different: can a person who is not you follow the runbook, get a working system back, and say how long it took? If the answer is no, the policy is a hope. I said the short version of that in the Cyber Essentials walkthrough. This is the drill.
What you are proving#
Pick numbers before you start, even rough ones. How old can the copy be (the recovery point). How long may the restore take before the business is harmed (the time you are about to measure, rather than the time on a slide). Who is allowed to do it.
Then prove four things, in this order:
- The recovery point exists and someone other than the backup admin can see it.
- The restore lands somewhere safe. A drill does not overwrite production.
- The thing that comes back actually runs: the VM boots, the database answers a query, the share mounts and a file opens.
- The time is written down, and every gap has a person next to it. "The network team" is not a person.
A drill that restores disks and then everyone goes to lunch has tested the vault. It has not tested the system.
One VM, one database, one file share#
Once a quarter is enough if you actually do it. Rotating the choice matters more than boiling the ocean. This quarter's VM should not be the same easy web server every time. Pick something the business would ask for: a line-of-business VM, a SQL Server, a department file share.
Use a Recovery Services vault as the baseline for those three when they are Azure Backup workloads. Write the date, the person, the stopwatch, and the link to the ticket in the same place every quarter. The log is the evidence. The screenshot of a green job is not.
The VM#
Restore the VM as a new virtual machine, into a resource group and a virtual network that are not production. Original-location overwrite is for the real incident, and only after someone has said the word.
Check the boring pieces the vault does not reconstruct for you:
- It boots, and someone can sign in.
- DNS for that test name points at it on purpose.
- NSGs and routes let the one admin in and do not let the test box talk to production databases.
- Extensions, managed identities, and anything that called a key vault still make sense, or you have written down that they do not.
- If the backup is application-consistent, say so. A crash-consistent copy of a busy database is a weaker drill. Prefer the application-consistent point when the workload has one.
File-level restore is useful, and it is a different test. Mounting a recovery point to copy one file does not tell you the VM boots. Do the file test when the ticket is "we deleted a spreadsheet". Do the boot test in the quarterly drill.
The database#
If the database is SQL Server on an Azure VM, restore with Azure Backup to a non-production instance. Then run a query a person in that team recognises: row count, latest order, a report that should match last Tuesday. Restoring the files and not starting the database is not the drill.
If the database is Azure SQL, it is probably not in that vault. Azure SQL has its own automated backups and point-in-time restore. Drill that on its own terms: restore to a new database, run the same kind of query, then delete the test copy. Ticking the vault box because a VM job is green, while the actual system of record is a PaaS database nobody has restored, is how the scope lies to you.
The file share#
Restore the Azure file share to an alternate location when the product allows it for that share. Mount it as a user who should have access, and as a user who should not. Open a file. Check the restore did not quietly land a second copy that everyone can now see.
Permissions and the path are the usual gap. The data is there, and the department still cannot reach it, because the drill never included the mount.
Make the backup harder to delete than the data#
Ransomware and a compromised admin both go for the recovery points. A vault your everyday Owner account can empty is part of the blast radius. A few controls change that, and the order matters.
Soft delete. On current vaults this is on, and in regions where Microsoft's secure-by-default assurance applies you cannot turn it off. Retention is 14 days unless you extend it, up to 180. Fourteen days is a short time if the person who would notice is waiting for the next quarterly drill. Lengthen it to a window you would actually catch. Soft delete is the undo for a deletion. It is not a substitute for the drill.
Immutability. An immutable vault blocks operations that would lose recovery points. You can lock that setting. The lock is irreversible, and it uses WORM storage so a later admin cannot switch immutability off and delete the copies. Run at least one restore while you can still change your mind, confirm the retention is the retention you meant, and then lock it. A locked ten-year retention you enabled by mistake is also locked.
A second person for the dangerous buttons. Multi-user authorization puts a Resource Guard in the path of destructive operations. Put that guard somewhere the backup admin does not already own. Test that a protected operation is actually refused. Microsoft is explicit that you should test this rather than assume the control is in force.
The admin of the vault. PIM if you have Entra ID P2, phishing-resistant MFA, a compliant device, no standing Owner used for daily mail. That is the same pattern as Conditional Access for admins and time-bound elevation. The vault is a privileged system. Treat the sign-in that way.
Geo-redundant storage and cross-region restore are a design choice you make before the region is the incident. Turn them on only for workloads that support them, and drill a restore in the paired region once, so the runbook is not theoretical. Availability zones keep you up inside a region. They are not a backup.
Alerts before you need them#
A failed job is the easy alert. Built-in backup alerts and an Azure Monitor action group will send it. Also alert on the job that did not run. A paused backup is quiet in a way a red X is not.
Alert on vault replication as well. If you are paying for geo-redundant storage and replication is unhealthy, you do not have the copy you think you have. Find that on a Tuesday.
Send the alerts to a place a person reads and to a ticket, not only to a shared inbox that has been full since 2022. For the first month after you wire them up, look weekly, the same way I look at sign-in logs after a Conditional Access change. An alert that fires and is ignored will still be ignored during the incident.
What the restore does not bring back#
The vault has the disks, the database, or the files. It does not have your Entra tenant, your Conditional Access, your DNS zone, or the runbook that was wrong about the subnet.
Write those dependencies in the drill notes when you trip over them. Each one is a gap with an owner: who fixes DNS, who confirms the identity the app uses, who tells the business the measured time. The owner updates the runbook before the next quarter. A gap with no owner is the same gap next quarter.
If the measured time is longer than the business can stand, that is a result, not a failed drill. Change the design (more frequent recovery points, a runbook that skips three meetings, a smaller system to restore first) or change the promise. Do not leave the slide saying one hour because the slide was written first.
A short checklist#
- Numbers written down: how old the copy may be, how long you are allowing, who runs it.
- One VM restored as a new VM into an isolated network. It boots. Someone signs in.
- One database restored to non-production and queried. PaaS databases drilled in their own backup system.
- One file share restored and mounted, including a user who should not have access.
- Stopwatch, person, and runbook corrections recorded.
- Soft delete retention long enough to notice. Immutability locked only after a successful restore.
- Destructive vault actions covered by a second authority, and tested.
- Alerts on failed jobs, missing jobs, and replication lag, going to a person.
Closing#
A backup policy is a schedule and a bill. A restore drill is the first time you find out whether the schedule was of any use.
Quarterly is not dramatic. One VM, one database, one file share, somewhere that is not production, with a clock running and a name next to every thing that broke. The green job can keep being green. You will already know whether you can put the system back.
References#
Sources and further reading for the claims above.
- Microsoft Learn, Azure Backup overview
- Microsoft Learn, Recovery Services vault overview
- Microsoft Learn, Restore Azure VMs
- Microsoft Learn, Restore files from an Azure VM backup
- Microsoft Learn, Restore SQL Server databases on Azure VMs
- Microsoft Learn, Restore Azure file shares
- Microsoft Learn, Azure SQL point-in-time restore
- Microsoft Learn, Soft delete for Azure Backup
- Microsoft Learn, Immutable vaults
- Microsoft Learn, Multi-user authorization
- Microsoft Learn, Monitor Azure Backup
- Microsoft Learn, Availability zones overview