Maintenance plan

Prepare an offline update as a complete release bundle

A device without dependable network access needs installation and rollback assets available before the update starts.

List every required asset

An offline release needs more than the application package. Identify the model, configuration, runtime image, conversion artifacts and device-specific prerequisites it requires. Include a manifest that names the intended hardware and software environment. Check the bundle on a representative spare device before deployment. A release that depends on retrieving one missing package at startup is not ready for a location where network access cannot be assumed, even if its main download has already been copied.

Preserve identifiable versions

Docker documents digest pinning as a way to retain the identity of a particular image despite mutable tags. Apply that principle to the broader release record: each component needs an identity that can be verified after transfer. Keep a readable release name for people and checksums or equivalent identifiers for files. Distinguish an intact copy from a compatible release. A checksum can establish that bytes match the manifest; the retained application checks still have to establish that those bytes work together.

Define promotion and rollback

Install the candidate without destroying the only accepted configuration. Specify the health checks required before it becomes active and the conditions that trigger rollback. systemd’s startup and restart semantics can help supervise the application, but they do not decide whether a new model produces acceptable results. Keep that acceptance logic in the release procedure. Confirm that rollback also restores compatible configuration and data expectations, rather than merely launching an older executable against whatever state the candidate left behind.

Carry the recovery instructions with it

Put the installation sequence, expected observations and fallback steps inside the bundle in a format available without the service running. Include the method for checking the active release after reboot. Rehearse a failed candidate and the return to service using spare storage or a disposable replica. The result should be a bounded maintenance operation with a known endpoint. This gives remote or infrequently visited devices a practical update path without depending on the memory of the developer who assembled the release.

Sources & publication notes

Retrospective period: September 2026. Written and published November 2, 2026. Product announcements establish release timing; support pages are linked as current references.

Related reading

← More guides & analysisCompare hardware →