[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-kill-switch::en":3,"gloss-cluster-kill-switch::en":26,"gloss-next-kill-switch::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"kill-switch","dev-tools","Kill Switch","A kill switch is a control that disables a feature immediately, without a deployment, a rebuild or a code change. It is usually a flag read at request time, and its entire purpose is to shorten the distance between noticing that something is wrong and stopping it. Everything else in an incident — diagnosis, root cause, the fix — can happen after the bleeding stops, and a kill switch is what lets it.\n\nAI features earn one more than most. Their failures are not the crashes that monitoring catches; they are behavioural, and behaviour can degrade because of an upstream model change, a prompt edit, a shift in the kind of input users are sending, or a retrieval index that silently went stale. None of those raise an exception. The system stays up, keeps responding, and produces answers that are wrong or inappropriate at a rate nobody notices immediately, which is precisely the situation where the ability to switch off in seconds is worth more than a fast rollback.\n\nA switch is only as good as what it falls back to. Off should mean something the product can survive: the previous non-AI behaviour, a simpler deterministic path, a queue that routes the work to a person, or an honest message explaining that a feature is temporarily unavailable. A flag that turns a feature into a blank screen or an error page will not be flipped when it matters, because the person holding it will weigh a visible outage against an invisible quality problem and choose to wait.\n\nGranularity is worth thinking about before the incident rather than during it. One switch per feature is the minimum. Better designs also allow disabling a single tool an agent can call, a single automated action, a single customer or region, or one step in a pipeline while the rest continues. The finest-grained switch that still stops the problem is the one to reach for, because a coarse switch turns a narrow failure into a broad withdrawal of service.\n\nThe operational details are what make it real. Who is allowed to flip it, and can they do so at three in the morning without waking anyone else? How fast does the change propagate — a flag cached for minutes is not a kill switch. Is the flip logged with a reason and an owner? Does it fail safe if the flag service itself is unreachable? And critically: has anyone tested it recently, in production, on purpose? An untested switch is a belief, not a control, and incidents are a bad time to discover the difference.\n\nThe last discipline is cleanup. Kill switches accumulate, and a codebase full of stale flags becomes hard to reason about, with dead branches nobody has executed in a year and combinations nobody has tested. Give each switch an owner and a review date, and remove the ones whose feature is now simply how the product works.","A kill switch turns a feature off in seconds without a deploy. For AI features it is the control that decides whether a bad day lasts minutes or a release cycle.",null,[11,14,17,20,23],{"slug":12,"name":13},"circuit-breaker","Circuit Breaker",{"slug":15,"name":16},"feature-flag","Feature Flag",{"slug":18,"name":19},"guardrails","Guardrails",{"slug":21,"name":22},"incident-management","Incident Management",{"slug":24,"name":25},"rollback","Rollback",[27,31,34,38,41,44,47,50,53,56,57,60],{"slug":28,"category":5,"name":29,"updated_at":30},"agent","Agent","2026-08-24T02:46:36+00:00",{"slug":32,"category":5,"name":33,"updated_at":30},"ai-code-assistant","AI Coding Assistant",{"slug":35,"category":5,"name":36,"updated_at":37},"api-gateway","API Gateway","2026-08-24T02:46:37+00:00",{"slug":39,"category":5,"name":40,"updated_at":37},"api-versioning","API Versioning",{"slug":42,"category":5,"name":43,"updated_at":30},"autonomous-agent","Autonomous Agent",{"slug":45,"category":5,"name":46,"updated_at":37},"blue-green-deployment","Blue-Green Deployment",{"slug":48,"category":5,"name":49,"updated_at":37},"canary-deployment","Canary Deployment",{"slug":51,"category":5,"name":52,"updated_at":37},"chaos-engineering","Chaos Engineering",{"slug":54,"category":5,"name":55,"updated_at":30},"ci-cd","Continuous Integration \u002F Continuous Deployment (CI\u002FCD)",{"slug":12,"category":5,"name":13,"updated_at":37},{"slug":58,"category":5,"name":59,"updated_at":37},"cli","Command-Line Interface (CLI)",{"slug":61,"category":5,"name":62,"updated_at":37},"cloud-development-environment","Cloud Development Environment (CDE)"]