Pick the property blue green prompt deployment gives you that a canary rollout does not
Blue-green keeps both prompt versions live at once and rolls back with a single atomic pointer flip; canary ramps traffic gradually so rollback means unwinding the ramp.
Think of a restaurant with two identical kitchens. Blue-green is when both kitchens are fully staffed and cooking, and a single switch decides which one's plates go to customers. If the new chef's menu is bad, you flip the switch back and the old kitchen instantly resumes service. Canary is more like sending the new chef's plates to one table first, then five tables, then twenty-five. If something goes wrong halfway, you have to walk around and pull plates from every table that already got them. Blue-green gives you that one instant undo. Canary gives you gentler exposure but a slower retreat. They solve different parts of the rollout problem.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Rollout patterns from classic web ops carry over to LLMOps, but the unit of change is a prompt rather than a container, and the cost structure is upside-down. Doubling your infrastructure to keep a second container warm used to be expensive. Doubling your prompt registry entries to keep a second prompt version warm is effectively free. That economic flip is why blue-green is the default rollout pattern for prompt changes in 2026, and it is also why an interview question that tests whether a candidate confuses blue-green with canary is testing whether they have actually shipped a prompt to production.
This explanation walks through what each pattern is, what the MCQ distractor options are testing, and how the two patterns combine in mature stacks.
What blue-green prompt deployment actually looks like
In the blue-green model, the prompt registry holds the current production prompt (blue) and the candidate new prompt (green) as two independently resolvable versions. The production pointer (often a label like summarize@production or summarize@live) names which version the runtime resolves at request time.
The deploy is a single pointer write that swaps the label from blue to green. The rollback is the inverse pointer write. Both operations are atomic at the registry level, complete in milliseconds at the registry, and propagate to the runtime as fast as the runtime's cache TTL allows (typically a few seconds on registries built for this).
The two important properties are simultaneity and atomicity. Simultaneity means both versions stay warm and resolvable, so the rollback does not need a rebuild, redeploy, or warmup. Atomicity means the flip happens in one transaction, so you never get a partial state where some requests resolve blue and others resolve green. Take either property away and the value proposition collapses.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- LangSmith Hub pointer flips: a single PATCH on the prompt's production tag rolls a regression back in under a second.
- Vercel AI Gateway's prompt registry exposes blue-green semantics natively; the production pointer is a first-class entity with atomic update.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you combine blue-green with canary so you keep both the slow-ramp safety and the fast-rollback property?
Promote the new prompt to green and route a percent of traffic to it via a feature flag (canary semantics). Once metrics hold across the ramp, atomic-flip the registry pointer so green becomes the new blue. Rollback during the ramp is a flag toggle; rollback after promotion is the inverse pointer flip.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Conflating canary and blue-green as interchangeable rollout strategies. They optimize for different goals: canary limits blast radius during the ramp, blue-green minimizes time to rollback after a regression is detected.
60 second bullets to scan on the way to the call.
What the blue and green environments represent in this pattern
Why the flip needs to be atomic and what breaks if it is not
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.