Sources checked . See the article’s scope for verification limits.
A rollback plan that says “restore backup” has left most of the interesting work for the outage.
Which backup? Onto what? With which credentials? What happens to data written after it was taken? A useful maintenance plan answers those questions while the system still works.
This guide is a planning and rehearsal workflow for a small environment you administer. It assumes you can identify the service's storage, obtain a supported backup, access its management console and use a separate recovery target. It does not provide a universal upgrade command. Product versions, database migrations and storage layouts deserve their own runbooks.
1. Name the change and the state it can affect
Start with a sentence: “Update the inventory application from the recorded current version to the selected supported version.” Then list the moving parts: application package or image, database, uploaded files, configuration and required credentials.
For a containerized service, record the persistent volumes and bind mounts as well as the image reference. Docker documents that volumes outlive individual containers. Our recommendation follows from that separation: do not call an image downgrade a complete rollback until you have accounted for the data it will open. Docker's storage documentation explains the lifecycle; it does not promise your application's older version will understand its newer database.
Read the application's release and migration notes before choosing a recovery method. If they do not document backward compatibility, treat that as an unresolved gate rather than assume the old binary can read the changed state.
2. Build a small change card
Here is an illustrative card for a fictional inventory service. The times are chosen planning limits, not measured performance or universal targets.
| Field | Example decision |
|---|---|
| Change | Application upgrade; database migration expected |
| Owner | Administrator performing the change |
| Scope | Application, database and uploads; no network redesign |
| Data-loss tolerance | No accepted user writes may be discarded |
| Maintenance window | 45 minutes; recovery decision by minute 15 |
| Before change | Stop user writes using the application's supported method; take the recovery set |
| Acceptance | Login, read a known record, create/read/edit a disposable record, retrieve an existing attachment |
| Stop condition | Migration error, missing attachment, or failed write check |
| Recovery | Keep writes stopped; restore the matched application/database/files set using the rehearsed procedure |
“No accepted writes lost” is a requirement to design around, not a feature this card creates. If the service cannot be made read-only or stopped safely, this example's recovery approach does not meet that requirement. Choose a different method or reschedule.
For a business service, add the approving owner and who tells users that access is restored. At home, tell anyone relying on the service. A smaller audience still appreciates finding out before the television stops working.
3. Match the backup to the application
A collection of files is not automatically a valid database backup. PostgreSQL's documentation distinguishes an ordinary file copy from a consistent filesystem snapshot: copying a running data directory with a general archive tool is not a supported substitute for consistency. A snapshot approach must cover the required data and write-ahead log coherently, including multiple filesystems where relevant. PostgreSQL's filesystem-backup guidance describes those constraints.
A logical PostgreSQL dump is another option. pg_dump produces an internally consistent view of one database; cluster-wide objects such as roles need separate handling. Restore also depends on required ownership and permissions. That is why a database dump belongs in a recovery set with the configuration and related files, rather than being treated as the whole service. Consult PostgreSQL's SQL-dump and restore guidance for the supported method.
Our recommendation: record the recovery set's capture point and whether application writes were paused. A database-consistent capture alone does not establish consistency with an attachment written to separate storage. Use the application's documented backup procedure to coordinate those parts.
Keep a recoverable copy outside the storage failure you are trying to survive. A convenient snapshot can be useful for undoing a change, but a snapshot still on the affected storage does not answer “what if this storage is unavailable?” State which failure your recovery set actually covers.
4. Rehearse without creating a second live service
Restore into a separate target with enough capacity. Before starting it, isolate it from production clients and disable outbound actions such as email, webhooks, scheduled jobs and synchronization. Use the product's supported controls; changing a hostname alone is not an isolation plan.
Check the exact restored copy: can you authenticate, read representative older data and retrieve an attachment? If appropriate, perform a disposable write and verify that it survives an application restart. Record what you checked and how long recovery took. Do not call a download verification or a successful archive listing a completed service restore.
If that exercise consumes more time than the maintenance window permits, change the window or recovery approach before proceeding. The rehearsal has already paid for itself.
5. Make one bounded change
Confirm console access and recovery credentials before beginning. Keep the old configuration and exact application version reference. Record a baseline of the same user actions you will test afterward.
Then follow the product's supported upgrade instructions, keeping the agreed write restriction in place. Avoid combining the application update with a storage migration and a firewall redesign. Our preference is one change whose effects you can recognize.
Use the prewritten checks, not whichever page happens to load first. Record failures even if a retry works; intermittent behavior is still evidence to evaluate. At the recovery deadline, choose explicitly whether to restore, investigate within an agreed extension, or stop. Do not spend the reserved recovery time hoping the progress indicator has a plan.
6. Recover the matched state, then reopen access
If acceptance fails, preserve relevant logs and the failed state when capacity allows. Recover the matched application, data and configuration set through the rehearsed method. Recheck the service before permitting writes.
Reverting old data after users resume writing may discard their newer work. That requires a separate reconciliation or forward-recovery decision; it is not a harmless retry of the original plan. Keep this boundary visible in the runbook.
Finally, note the actual result, new version, recovery evidence and remaining issues. Keep the recovery set according to an explicit retention decision. A maintenance job is finished when the required behavior is back and the next operator can understand the state left behind.
Sources and scope
Primary documentation checked 2026-11-17. This is documentary guidance with an original illustrative change card; no upgrade, backup or restore was executed for this article. Verify the installed product's version-specific procedure before use.