MCP Academy · 课时

安全失败，绝不挂起

防范超时和未处理的异常。

安全失败，绝不挂起 是 CoddyKit 上的免费 MCP Academy 课时。这是第 4 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 MCP Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 MCP 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。

A Hung Tool Is Worse Than a Failed One

A tool that errors clearly lets the model move on. A tool that hangs freezes the whole call, so always design for a timely exit.

Catch the Unexpected

Wrap risky work so a surprise exception becomes a clean error, not a crash. The server stays up and the model learns what failed.

try:
    data = parse(raw)
except Exception as e:
    raise ValueError("could not parse input")

Always Set a Timeout

Any network or external call needs a timeout. Without one, a slow upstream can pin your tool open indefinitely and stall the client.

import httpx
r = httpx.get(url, timeout=10.0)

Prefer Async for I/O

Async tools let one slow request wait without blocking everything else, so your server keeps serving other calls meanwhile.

@mcp.tool()
async def fetch(url: str) -> str:
    async with httpx.AsyncClient() as c:
        r = await c.get(url, timeout=10)

Honor Cancellation

If the client cancels, stop promptly. Respecting cancellation frees resources and avoids work nobody is waiting for anymore.

Bound Loops and Retries

Cap any loop or retry. An unbounded retry on a failing API can spin forever, so set a max attempts and then give up gracefully.

for attempt in range(3):
    if try_call():
        break

Release What You Open

Use with blocks or finally so files and connections always close, even when an error fires partway through the function.

with open(path) as f:
    return f.read()

Degrade Gracefully

When part of a job fails, return what you have with a note rather than nothing. A partial result still helps the model decide.

Log the Failure Server-Side

Record the full detail in your logs while sending the model only a safe summary. You get to debug without leaking internals.

One Crash Shouldn't End the Server

A single bad call must not take down the process. Isolating failures per tool keeps every other capability available to clients.

Always Return Something

Every path through your tool should end in a response, success or error. Silence leaves the client waiting and the model stuck.

Quick Check

What is the key reason to add a timeout to an external call in a tool?

Recap

Fail fast, never hang: catch surprises, set timeouts, bound retries, release resources, log safely, and always return a response. ✅

常见问题解答

「安全失败，绝不挂起」课时是免费的吗？

是的 — 「安全失败，绝不挂起」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 MCP Academy 课程的其余内容，请升级到 CoddyKit PRO。 MCP Academy 课程共包含 4 节课。

「安全失败，绝不挂起」这节课中我会学到什么？

防范超时和未处理的异常。你通过在浏览器中直接运行的动手代码来练习 MCP 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 MCP Academy 需要有经验吗？

无需任何先前经验。CoddyKit 上的 MCP 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 4 节课，共 4 节。

「安全失败，绝不挂起」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 MCP Academy 课中编写并运行代码吗？

能。每节 MCP 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 MCP Academy