Abstract
A method is disclosed for letting the outcome of a post-deploy smoke test decide which of two independent mechanisms holds rollback authority over a newly deployed image. The smoke gate runs an ordered list of checks against the new instance while traffic is held on a maintenance page. The checks are partitioned by position into a core section establishing whether the image boots (liveness, startup, readiness with the database up, session, single-sign-on discovery) and an auxiliary section intended to exercise features through an instance that boots (a language-model round trip, a plug-in load report that must answer healthy and, among failed plug-ins, fails only on those absent from a baseline, a feature-router mount check). No check short-circuits the others, each failure counts against the section in force, and the gate returns one of four statuses: pass, core failure, auxiliary-only failure or skipped. A deploy driver maps the status onto a cluster-resident authority record read by an independent reconciler that, in its default mode, restores the authorized image when the running image is unauthorized and not ready. On a core failure the driver holds the maintenance page, pages an operator and leaves the previous image authorized, so the reconciler's rescue stays armed. On an auxiliary-only failure it also holds and pages, but writes the held image as authorized before the reconciler resumes, so working code cannot be reverted at a later unrelated readiness blip and rollback becomes a human decision. An enforced pre-traffic provenance check always takes the core mapping.
Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Assuncao, gustavo matthew, "Failure-Class-Partitioned Post-Deploy Smoke Gate Whose Boots-Versus-Works Verdict Decides Whether an Independent Reconciler's Automated Rollback Stays Armed or Passes to a Human", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/13203