MLOps Academy · 课时

缩容到零并恢复运行

通过请求驱动的自动扩缩容节省成本

缩容到零并恢复运行 是 CoddyKit 上的免费 MLOps Academy 课时。这是第 2 节课，共 4 节。你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。这是 MLOps Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 MLOps Academy 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

Idle Models Cost Money

A model pod sitting with no traffic still burns CPU, memory, and cloud bills. KServe can shrink an idle service all the way down. Scale to zero stops that waste. 💸

Powered by Knative

KServe's serverless mode rides on Knative, which watches request volume and adjusts replicas. When traffic stops, it can remove every pod for that service.

Request-Driven Autoscaling

Replicas track demand, not a fixed schedule. As requests rise, KServe adds pods, and as they fade it scales back. This is request-driven autoscaling in action.

Setting minReplicas to Zero

To allow full scale down, you set minReplicas to 0 in the predictor spec. With zero allowed, an idle service drops to no running pods at all.

spec:
  predictor:
    minReplicas: 0
    model:
      modelFormat:
        name: sklearn

The Concurrency Target

The autoscaler aims for a target number of in-flight requests per pod. This concurrency target decides how aggressively KServe adds replicas under load.

metadata:
  annotations:
    autoscaling.knative.dev/target: "10"

Scaling Back Up

When a request arrives at a zero-scaled service, Knative spins a pod up on demand. Traffic resumes the moment the pod is ready, so the scale up is automatic.

The Cold Start Cost

That first request after zero must wait for the pod to start and load the model. This delay is the cold start, the main trade-off of scaling to zero. ⏱️

When Zero Makes Sense

Scale to zero shines for bursty or rare workloads where idle time dominates. For steady, latency-critical traffic, the cold start penalty may not be worth it.

Keep One Warm Instead

If cold starts hurt, set minReplicas to 1 so one pod always stays alive. You trade a little cost for a guaranteed warm instance ready to serve.

spec:
  predictor:
    minReplicas: 1

Cap the Upper End

You can also bound growth with maxReplicas so a traffic spike never overruns your cluster budget. It puts a ceiling on the autoscaler.

spec:
  predictor:
    minReplicas: 0
    maxReplicas: 5

Watch It Scale

You can confirm the behavior by watching pods appear and vanish with traffic. The pod count drops to zero when idle and climbs back when calls come in.

kubectl get pods -l serving.kserve.io/inferenceservice=sklearn-iris -w

Quick Check

What is the main downside of letting a service scale to zero?

Recap

You saw how minReplicas: 0 lets KServe scale idle models to zero and back on demand. Cap with maxReplicas, and keep one warm if cold starts hurt. Great progress! 🎉

常见问题解答

「缩容到零并恢复运行」课时是免费的吗？

是的 — 「缩容到零并恢复运行」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 MLOps Academy 课程的其余内容，请升级到 CoddyKit PRO。 MLOps Academy 课程共包含 4 节课。

「缩容到零并恢复运行」这节课中我会学到什么？

通过请求驱动的自动扩缩容节省成本你通过在浏览器中直接运行的动手代码来练习 MLOps Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 MLOps Academy 需要有经验吗？

无需任何先前经验。CoddyKit 上的 MLOps Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。这是第 2 节课，共 4 节。

「缩容到零并恢复运行」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 MLOps Academy 课中编写并运行代码吗？

能。每节 MLOps Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

← 返回 MLOps Academy