MLOps Academy · 课时

在不同模型版本之间分配流量

将一定比例的请求路由到挑战模型

在不同模型版本之间分配流量 是 CoddyKit 上的免费 MLOps Academy 课时。这是第 1 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。

Two Models, One Live Test

You have a new model you believe is better. Instead of guessing, you let real traffic decide. An A/B test runs both models side by side on live users. 🔬

Champion and Challenger

The current model in production is your champion. The new candidate you want to prove is the challenger. The test sees if the challenger truly beats the champion.

Splitting the Traffic

The core idea is simple: send a slice of requests to each model. A common start is 90/10, keeping most users safe on the champion while the challenger proves itself.

A Tiny Router

A basic split just rolls a random number per request and routes by the cutoff. Here ten percent of traffic reaches the challenger.

import random

def pick_model(challenger_share=0.1):
    return "challenger" if random.random() < challenger_share else "champion"

Keep Each User Consistent

Random per request flips a user between models on every visit. Instead, hash the user id so the same person always lands on the same model.

Hashing for Sticky Splits

Hashing the user id gives a stable bucket. The same id maps to the same model every time, which keeps the test clean.

import hashlib

def bucket(user_id, challenger_share=0.1):
    h = int(hashlib.md5(user_id.encode()).hexdigest(), 16)
    return "challenger" if (h % 100) < challenger_share * 100 else "champion"

Log Which Model Served

For every prediction, record which model handled it. Without this assignment log, you can never compare the two groups fairly later on. 📝

Start Small, Then Ramp

Begin with a tiny challenger share to limit risk. As confidence grows, you raise the percentage gradually instead of flipping everyone over at once.

Control the Split with Config

Hard-coding the share means a redeploy to change it. Read the split ratio from config so you can dial traffic up or down without shipping code.

Same Inputs, Fair Fight

Both models must see the same kind of requests and the same features. If the groups differ in who they serve, any winner you find may be an illusion.

It Is Just Routing

At its heart, an A/B test is a routing layer plus careful logging. Get the split sticky and recorded, and you have the foundation for a trustworthy comparison. ✅

Quick Check

Let us check how to keep your split clean.

Recap

An A/B test pits a champion against a challenger by splitting live traffic. Hash users for sticky buckets, log every assignment, and start small. 🎯

常见问题解答

「在不同模型版本之间分配流量」课时是免费的吗？

是的 — 「在不同模型版本之间分配流量」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。

「在不同模型版本之间分配流量」这节课中我会学到什么？

将一定比例的请求路由到挑战模型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 1 节课，共 4 节。

「在不同模型版本之间分配流量」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