
Open Source & Linux · Cloud & Infrastructure · ~13 min read
Switching a server on takes an afternoon. Switching one off properly takes about three weeks, touches a dozen systems and is where most infrastructure documentation quietly goes wrong. We collected the questions readers send us most about retiring servers — physical and cloud — and put them to Puffin Notes editor Owen Calder.
Q&A WITH OWEN CALDER · NOVEMBER 2026 · CHECKLIST INCLUDED
$ cat summary.txt
7
Phases between “we don’t need this box” and a closed change record
2–3 wks
Typical end-to-end time to back up, remove and replace a server
14+
Systems that usually hold a reference to any given server
$118/mo
What a “deleted” AWS server can keep costing through leftovers (our calculation)
// part one · before you touch anything
The decision, the paperwork and the backup
Why does something as simple as switching off a server need a whole process?
Because it isn’t one action, it’s a chain — and every link is somewhere a different team keeps records. The server is in your monitoring, your backups, your DNS, your firewall rules, your config management, probably three wikis and at least one spreadsheet someone made in 2019. Powering it down only changes one of those. Everything else now describes a machine that doesn’t exist.
The informal version — someone who knows the box shuts it down, a colleague helps carry it out, the docs get updated “later” — works right up until it doesn’t. Then you get a monitoring alert for a host nobody can find, a restore from a backup that turns out to be six months old, or a new server handed an IP address that an old cron job is still trying to reach.
IT provider Extech Cloud puts the realistic timeline at two to three weeks to back up, remove and replace a server. That matches my experience. Most of that time isn’t work; it’s waiting to be sure.
What’s the very first step?
Paperwork, and I mean that seriously. Before anything is powered down, get written approval from the application owner — the one person who can say with authority that nothing important still runs there — and open a change record that says which server, why, when, and who approved it. Six months later, when an auditor asks where a device went, that record is your answer. Microsoft’s guidance on decommissioning source workloads after a migration makes the same point: stakeholder sign-off, reclaimed licences and archived data come before anything is switched off.
Then write down the server’s side jobs. The headline application is never the whole story. Old servers pick up roles over the years: a log forwarder, a backup client, a local DNS cache, a monitoring exporter, the cron job that rotates someone else’s certificates. Each one either moves or gets switched off deliberately — otherwise you find out when a dashboard goes dark.
# who is talking to it right now? $ ss -tunap | grep ESTAB # scheduled work, for every user $ for u in $(cut -d: -f1 /etc/passwd); do crontab -l -u "$u" 2>/dev/null; done $ systemctl list-timers --all # what is listening, and who logged in lately $ ss -tlnp $ last -n 50
How do I know my backup is good enough?
You restore it. That’s the whole answer, and it’s the step almost everyone skips. A decommissioning backup that captured the app directory but not the database, or the data but not the encryption keys, is worse than useless — it gives you confidence you haven’t earned.
Three rules. Restore it on a different machine. Store it somewhere that survives the server — not a share on the box you’re about to wipe. And write the retention period and owner into the change record, so the data gets deleted deliberately when its time is up instead of lingering forever.
$ cat timeline.txt
What three weeks actually looks like
Puffin Notes model for a single physical server with one main application. Active work is a few days; most of the calendar is observation, which is the point. A cloud instance skips phases 5 and 6 but not the rest.
// part two · cutting the strings
Deregister first. Shut down second.
Where do references to a server actually live?
More places than anyone remembers. This is the list I work through. If you only take one table from this post, take this one.
| System | What to remove or change | If you forget |
|---|---|---|
| DNS | A, AAAA, CNAME and PTR records | Names resolve to an IP that may be reassigned |
| DHCP / IPAM | Reservations; release or re-reserve addresses | Address conflicts when the IP is reused |
| Monitoring & alerting | Hosts, checks, dashboards | Alerts for a ghost — and a team that learns to ignore alerts |
| Backup schedules | Jobs and agents | Nightly failures that hide real failures |
| Config management | Inventory entries (Ansible, Puppet, Salt) | Playbooks that fail or time out on every run |
| Firewall & security groups | Rules referencing its IPs | Open paths to an address someone else may get |
| Load balancer pools | Backend members | Health checks failing, traffic routed to nothing |
| Directory (AD / LDAP) | Machine and service accounts | Live credentials for a machine that’s gone |
| TLS certificates | Revoke and record | A valid credential that outlives the host |
| SSH trust | Keys in other hosts’ authorized_keys, known_hosts | Old keys that still open doors elsewhere |
| CI/CD deploy targets | Inventories and pipeline variables | Failed deploys, or deploys to a recycled IP |
| Licences | Release seats and keys | Paying for software on a dead machine |
| CMDB / asset register | Status to “retired” with disposal details | An inventory nobody trusts |
| Runbooks & wikis | Mark retired, point to the replacement | 3 a.m. instructions for a server that isn’t there |
Puffin Notes list, building on the deregistration steps in Obelinf’s server decommissioning checklist.
What’s a scream test, and is it a real thing?
Very real, slightly unprofessional name. You power the server off — but leave it racked, wired and intact — for a fixed observation window, and you wait to see who screams. Any unexpected alert, failed job or confused colleague in that window is a dependency you missed in step four. Finding it while you can still press the power button is cheap. Finding it after the drives have been wiped is not.
I use a week as a minimum, because it catches every daily job and most weekly ones. If the server ran anything monthly — billing exports, reports, certificate renewals — either extend the window past the next run or trigger that job manually while the server is off. Use the window for the cleanup that was too risky while the machine was live: disable its service accounts, remove the machine from the domain, revoke its certificates.
// part three · making the data actually go away
Wipe it before it leaves the building
Isn’t deleting the partitions enough?
No. Deleting partitions or reformatting removes the map, not the data — anyone with free recovery tools can read most of it back. Once a drive leaves your premises for recycling, resale or a vendor return, you’ve lost control of what’s on it, so erasure happens while the server is still yours, and gets verified and written down.
The right method depends on the media. The widely used reference here is NIST’s media sanitisation guidance (SP 800-88), which sorts methods into three levels: clear, purge and destroy. The important nuance is that SSDs and spinning disks behave very differently.
| Media | Recommended method | Linux tooling | Watch out for |
|---|---|---|---|
| Spinning disk (HDD) | Full overwrite, or ATA Secure Erase; shred if not reused | shred, hdparm | Slow on large disks — budget hours, not minutes |
| SATA SSD | Firmware secure erase or cryptographic erase | hdparm –security-erase | Overwriting is unreliable on SSDs because of wear levelling |
| NVMe SSD | Format with secure erase, or sanitize | nvme format, nvme sanitize | Check the drive supports the mode you choose |
| Self-encrypting drive | Cryptographic erase (destroy the key) | vendor or sedutil tools | Only as good as the drive’s implementation |
| Drives behind a RAID controller | Erase via the controller, or pass drives through first | controller CLI | The OS may not see the physical disks at all |
| Firmware & management controller | Factory-reset BMC/iLO/iDRAC, clear BIOS, remove USB/SD boot media | ipmitool, vendor UI | Cached passwords and network config survive disk wipes |
Summary based on NIST SP 800-88 categories. Commands erase data irreversibly — double-check the device name before you press enter.
# confirm which device is which, twice $ lsblk -o NAME,SIZE,MODEL,SERIAL # NVMe: user-data erase (1) or cryptographic erase (2) $ sudo nvme format /dev/nvme0n1 --ses=1 # HDD: one random pass, then zeros $ sudo shred -v -n 1 -z /dev/sdb # verify: this should print nothing but zeros $ sudo head -c 1M /dev/sdb | hexdump -C | head
Then record the method, the date, the drive serial numbers and who did it in the asset register. That line is what makes your disposal defensible in an audit. If you use a certified disposal vendor, keep their certificate with the record.
// part four · the cloud version
You can’t unrack a VM. You can still get it wrong.
Isn’t decommissioning a cloud server just clicking “terminate”?
That’s exactly the problem. Terminate removes the instance and leaves its surroundings behind: detached volumes, snapshots, the elastic IP, the load balancer that only existed for it, the NAT gateway in a subnet nothing uses any more. None of them make noise. All of them keep billing.
I priced a fairly typical leftover set on AWS at public list prices. The server itself is gone. The bill isn’t.
So what’s the cloud checklist?
Everything in part one and two still applies — approval, backup, deregistration, scream test. Stopping an instance is the cloud’s version of leaving it racked but powered off. Then, before you call it done:
| Step | Physical server | Cloud instance |
|---|---|---|
| “Off but recoverable” | Powered down, still racked | Stopped (storage still billed) |
| Data erasure | Your job, on premises | Delete volumes and snapshots; the provider handles media |
| Network identity | Release IPs in IPAM, clean switch ports | Release elastic IPs, delete unused ENIs |
| Surrounding infrastructure | Cables, PDUs, rack units | Load balancers, NAT gateways, security groups |
| The hidden cost | Power, cooling, rack space | Orphaned resources billing silently |
| Proof it’s done | Disposal certificate, asset register | Empty resource list for its tags; cost report showing zero |
The single most useful habit: tag everything with the service it belongs to when you create it. Then “what’s left of this server?” becomes one filtered query instead of an afternoon of clicking.
# volumes not attached to anything $ aws ec2 describe-volumes --filters Name=status,Values=available # elastic IPs with no association $ aws ec2 describe-addresses --query "Addresses[?AssociationId==null]" # everything still carrying the old service tag $ aws resourcegroupstaggingapi get-resources --tag-filters Key=service,Values=legacy-api
$ git commit -m
“A deliberate retirement leaves a trail of closed tickets. An accidental disappearance leaves a trail of confusion.”
— Owen Calder
// part five · after it’s gone
The step everyone skips
The hardware’s gone. Are we done?
Almost, and this is where most teams stop one step short. Reconciling the documentation — IPAM, VLANs, topology diagrams, the CMDB, runbooks, disaster recovery plans — is the longest and least visible part of the job, which is exactly why it gets deferred forever. Make it a checklist item with a date, not an intention.
When you pull a server, also note what you found that the documentation didn’t mention: the extra network card, the second power supply, the label that doesn’t match the hostname. Those discrepancies are the best free audit you’ll ever get of how accurate the rest of your records are.
Finally, don’t close the change record the day the hardware leaves. Give it a settling week, check that nothing surfaced, then close it with a summary: what was done, where the backup lives and for how long, and where the hardware went.
What’s the one mistake you see most often?
Releasing the IP address too early. Someone frees it in IPAM the day the server goes off, a new machine picks it up a week later, and suddenly an old firewall rule, a forgotten cron job or a monitoring check is pointed at a completely different server. Sometimes that’s merely confusing. Sometimes it’s a security hole. Hold the address until the record is closed.
$ cat decommission.md
The full checklist
## 1 · approve
- Written sign-off from app owner
- Change record opened
- Side jobs listed and reassigned
- BMC credentials located
## 2 · back up
- Data, DB, config and keys captured
- Restore tested on another machine
- Retention and owner recorded
## 3 · deregister
- DNS, DHCP, monitoring, backups
- Config mgmt, LB pools, CI targets
- Firewall rules reviewed
## 4 · shut down
- Off in a maintenance window
- 7+ day scream test passed
- Accounts disabled, certs revoked
## 5–6 · erase & remove
- Drives erased and verified
- BMC and BIOS reset
- Cables, rack units, disposal recorded
## 7 · reconcile
- IPAM, CMDB, diagrams, runbooks updated
- IPs released only now
- Settling week passed, record closed
$ man decommission
Quick answers
How long does decommissioning a server take?
Around two to three weeks end to end for a typical server, though the hands-on work is only a few days. Most of the time is the observation window and the final settling period.
What’s the difference between decommissioning and disposal?
Decommissioning removes a server from service and from your records. Disposal is what happens to the hardware afterwards — reuse, resale, recycling or destruction. Disposal should only start once the drives are verifiably erased.
How long should I keep the final backup?
As long as your legal, contractual or compliance obligations require for that data — and no longer. Write the date into the change record so the deletion is deliberate.
Is formatting a drive enough before recycling?
No. Use a secure erase or cryptographic erase for SSDs, a full overwrite or secure erase for spinning disks, and verify the result. Physical destruction is the fallback.
What is a scream test?
Switching a server off, but leaving it intact, for a fixed period to see whether anything breaks. It reveals dependencies you missed while the change is still easy to undo.
What do cloud servers leave behind?
Most often unattached volumes, snapshots, elastic IPs, load balancers and NAT gateways. Terminating the instance doesn’t delete them, and they keep billing.
Sources: André Ribeiro, “The Complete Server Decommissioning Checklist”, Obelinf, August 2026; Extech Cloud, “Everything you need to know about server decommissioning”; NIST SP 800-88 media sanitisation categories; AWS published list prices for the leftovers calculation, which is our own and illustrative. Commands are examples — test them somewhere safe first.
