News & Entertainment for Mason City, Clear Lake & the Entire North Iowa Region
DON’T MISS NIT’S SUBSCRIBER-ONLY STORIESExclusive North Iowa reporting · Just $24 for 6 monthsUNLOCK STORIES
News Archives
Image Sets Decide How Long an Upgrade Window Runs
Published on
Facebook
Tumblr
Threads
X
LinkedIn
Email
An upgrade window is usually estimated from the application image alone. In a FastGPT self-hosted deployment, the application image is one of several that move together, and the window length follows from the set.
The 4.17.3 release notes list the component tags: the main service and the commercial edition at v4.17.3, the plugin at v1.2.0, the code sandbox at v5.0.0, the sandbox proxy, agent sandbox and volume manager at v0.4.0, and AI Proxy at v0.7.7. Recording all of them in the change ticket takes a minute and prevents a rollback by memory.
Why the set matters
Agent sandbox previews depend on compatible application, proxy and sandbox releases. The v4.17.3 notes require the preview proxy to be updated to v0.4.0 when this capability is enabled. Check the release requirements for the components actually installed, then test previews through the application. Keep this compatibility check separate from the independently versioned code execution sandbox, whose image follows its own release schedule.
That constraint turns an upgrade into three phases rather than one: replace and start the image set, let the post-upgrade tasks run, then verify. Each phase has its own duration, and only the first is under direct operator control.
Three phases, three measurements
Record four timestamps: upgrade start, images ready, backfill complete, regression passed. The three intervals between them are the data needed for the next estimate.
Backfill is the phase most often underestimated. After start-up dependencies complete, three background tasks run, sharing up to three concurrent slots: sandbox instance team ownership, folder ancestor paths, and system model price migration. Large datasets lengthen the queue.
Price migration needs its own check
System models without migrated pricing are treated as free in this version. Totals will not match usage until the migration is confirmed, so pricing verification belongs alongside application regression in the post-upgrade checklist.
Check that system models carry the new pricing structure, then compare one invoice against manually calculated usage. The same check belongs in the quarterly routine, because team models are excluded from migration and stay free by design.
Regression scope follows the integration surface
A deployment with many integrated interfaces needs a longer regression phase. Run success and failure paths for read, write and approval tiers, plus one identity check, and record results against the interface inventory.
A rollback plan states four things: which image set the deployment returns to, whether the data remains compatible, whether backfill tasks need reversing, and how the rollback is verified.
Data compatibility is the part that bites. A structure written by the new version may not be readable by the previous one, so a rollback taken without that check can convert a degraded state into an unusable one. Rehearse the rollback with the same care as the upgrade, and record both durations in the change log.
Distinguish the Agent sandbox and preview proxy from the independently released code execution sandbox. Recovery follows the compatibility requirements for the affected components; check data and configuration compatibility as well as image tags before choosing the rollback sequence.
Communicate the window, not just the schedule
A maintenance window needs two messages: one before it opens and one after it closes. The first states start and end time, the affected scope and the expected degradation. The second states the outcome and the backfill results.
Skipping the second message leaves users guessing whether the system is finished or merely running. That guess produces support tickets that cost more than the message.
Keep both messages in the change record. The next window can reuse the wording, and the estimated duration stops being a fresh negotiation each time.
Keep a rollback pair in the change log
Two tags belong in every change record: the set deployed and the set replaced. With several images moving together, the second is the one that gets lost, and losing it turns a routine rollback into a reconstruction exercise.
Store the pair next to the backfill results. Include the configuration revision, backup reference and operator for each component, so recovery preparation does not depend on which engineer performed the original installation. A single page then answers what changed, what ran afterwards, and what the deployment would return to.
Keeping that page for four releases produces a pattern: which phase dominates, and whether the estimate should be revised.