Skip to content

README.md

Latest commit

 

History

History
198 lines (154 loc) · 10.5 KB

README.md

File metadata and controls

198 lines (154 loc) · 10.5 KB
Raw

Dolphin MCP Pilot — Community Case Studies

A gallery of real-world usage stories: people driving Apache DolphinScheduler through an AI agent via dolphin-mcp-pilot, written up here and shared publicly (Xiaohongshu / RED, X, 掘金, CSDN, 知乎, Bilibili, YouTube, a personal blog, …).

Where examples/ collects reusable client configurations, cases/ collects what people actually did with them — the scenario, the conversation, the before/after, and a link to the public post that tells the story.

中文说明见下方

What a good case shows

A case is not a feature list — it is a short, honest story of a task you got done. The strongest cases share three things:

  1. A concrete task, not a demo for its own sake. "Our nightly ETL failed at 3am and I had the agent find the failed instances and rerun-from-failure from my phone" beats "I listed projects".
  2. The agent actually operating DolphinScheduler — a screenshot or short recording where you can see the natural-language request and the dolphin-mcp-pilot tool call / result. Read-only listing is fine as a supporting shot, but the interesting cases change state (create a workflow, fix a schedule, rerun a failed run).
  3. A public post other people can read, with the link recorded in your case file.

Creative directions

Pick one angle (or combine a few). These map to what dolphin-mcp-pilot is actually good at — see the tool list in the main README:

  • 自然语言运维 / NL ops — operate DolphinScheduler in plain language from an MCP host (Claude Code, OpenClaw, Cursor, CodeBuddy, Claude Desktop). "Pause the sales-pipeline schedule, I'm doing maintenance."
  • Incident firefighting — an instance failed; use the agent to read task logs, force a task success / skip a stuck task, or rerun-from-failure — end to end, without opening the DS console.
  • One-line workflow creation — turn a sentence or a SQL snippet into a real SQL / DAG workflow, then bring it online.
  • Schedule management by chat — create / online / offline / delete cron schedules conversationally, e.g. "move the daily report to 7am on weekdays only".
  • Monitoring & morning-brief — have the agent summarize last night's runs, surface failures, and propose reruns.
  • Version safety — roll back a workflow version or clone a workflow before a risky change.
  • Multi-tenant / team — show per-request auth (X-DS-Token or X-DS-User/X-DS-Password) letting different teammates drive the same server with their own DolphinScheduler credentials.
  • Host integration write-up — a focused "how I wired dolphin-mcp-pilot into " walkthrough, with the gotchas you hit.
  • Before / after — the same task done by hand in the DS web console vs. driven by the agent; what got faster or safer.

Bonus points for cases that combine dolphin-mcp-pilot with other tools in a real pipeline, or that teach a non-obvious trick (e.g. using the raw-API passthrough as a safety valve).

Requirements

To keep the gallery trustworthy, every submission must meet all of these:

  1. Real usage. You actually ran dolphin-mcp-pilot against a DolphinScheduler instance. The case includes at least one screenshot or short recording showing the agent's request and the dolphin-mcp-pilot tool call / result. Mock-ups, staged screenshots, or unverified text do not qualify.
  2. A public post. The story is published on at least one public channel with an openly reachable URL (no login-walled or private links). Record every link in the frontmatter channels: list. The post should be substantive — a few sentences plus a visual, not a bare repost of this repo's README.
  3. Original content. Your own writing and screenshots. No plagiarism, no AI-generated text passed off as a firsthand account. If you used AI to help draft the post, that's fine — just keep the events real.
  4. No secrets. Redact tokens, passwords, X-DS-* header values, internal hostnames/IPs, and any private business data in every screenshot and config snippet. Use obvious placeholders.
  5. This case file links back. The README.md you submit here must link to the public post(s), so a reader can go from the gallery to the full story.

How to submit

  1. Fork this repository.

  2. Copy cases/TEMPLATE/ to cases/<your-case-id>/ (short, unique, kebab-case, e.g. nightly-etl-rerun).

  3. Fill in README.md — complete the frontmatter and every section, and paste your public link(s). Add preview.png (or more images) showing the agent operating DolphinScheduler.

  4. Open a Pull Request. Suggested title:

    [case] <case-id> — <one-line what you did>

Multiple cases? One directory per case, one PR is fine.

Directory layout

cases/
├── README.md                 # this file — directions and requirements
├── TEMPLATE/
│   └── README.md             # copy this to start a case
└── <your-case-id>/
    ├── README.md             # required — frontmatter + the story + public links
    ├── preview.png           # required — the agent operating DolphinScheduler
    └── *.png / *.gif         # optional — more screenshots

Frontmatter

Each case README.md starts with YAML frontmatter:

---
title: Rerunning a failed nightly ETL from my phone
author: your-github-handle
date: 2026-08-14                # YYYY-MM-DD, when you published
category: incident-firefighting # one of the creative directions above
host: claude-code               # the MCP host you used (claude-code / openclaw / cursor / codebuddy / claude-desktop / other)
testedWith: dolphin-mcp-pilot 0.2.0
channels:                       # one or more PUBLIC links to your published post(s)
  - https://www.xiaohongshu.com/...
  - https://x.com/...
---

Dolphin MCP Pilot — 社区使用案例

一个真实使用故事画廊：用户通过 dolphin-mcp-pilot、借助 AI 智能体来操作 Apache DolphinScheduler，把过程写在这里，并发布到公开渠道（小红书、X、掘金、CSDN、知乎、B 站、 YouTube、个人博客……）。

examples/ 收录的是可复用的客户端配置，cases/ 收录的是 大家真正用它做成了什么——场景、对话、前后对比，以及一条能读到完整故事的公开链接。

一个好案例长什么样

案例不是功能罗列，而是一个诚实的小故事：你用它把某件事做成了。最打动人的案例有三个共性：

  1. 一个具体任务，而非为演示而演示。"凌晨 3 点 ETL 挂了，我在手机上让智能体找到失败实例并 从失败处重跑"，比"我列了一下项目列表"有力得多。
  2. 智能体真的在操作 DolphinScheduler——截图或短录屏里能看到自然语言请求 + dolphin-mcp-pilot 的工具调用/结果。只读列举可以作为辅助画面，但真正有意思的案例会改变状态 （新建工作流、修调度、重跑失败实例）。
  3. 一条公开帖子，别人能读到，链接记录在你的案例文件里。

创作方向

任选一个角度（也可组合）。这些正好对应 dolphin-mcp-pilot 擅长的能力（见 主 README 的工具清单）：

  • 自然语言运维——在 MCP 宿主（Claude Code、OpenClaw、Cursor、CodeBuddy、Claude Desktop）里 用大白话操作 DolphinScheduler。"把销售管道的调度下线，我要做维护。"
  • 故障救火——某个实例失败了：让智能体读任务日志、强制任务成功 / 跳过卡住的任务、或从失败处 重跑，全程不打开 DS 控制台。
  • 一句话建工作流——把一句话或一段 SQL 变成真正的 SQL / DAG 工作流并上线。
  • 对话式调度管理——会话式地 创建 / 上线 / 下线 / 删除 cron 调度，例如"把日报挪到工作日早 7 点"。
  • 监控与晨报——让智能体总结昨晚的运行、暴露失败项、并建议重跑。
  • 版本安全——高风险改动前回滚工作流版本或克隆工作流。
  • 多租户 / 团队——展示逐请求鉴权（X-DS-Token 或 X-DS-User/X-DS-Password），让不同同事 用各自的 DolphinScheduler 凭据驱动同一个 server。
  • 宿主接入手记——一篇聚焦的"我如何把 dolphin-mcp-pilot 接进 <某 MCP 宿主>"，含踩过的坑。
  • 前后对比——同一个任务，DS 网页控制台手工做 vs 智能体驱动，快在哪、稳在哪。

加分项：把 dolphin-mcp-pilot 和真实流水线里的其他工具组合起来，或讲清一个不显而易见的技巧 （比如用 raw-API passthrough 当安全阀）。

要求

为了让画廊可信，每份提交必须同时满足：

  1. 真实使用。你确实对一个 DolphinScheduler 实例跑过 dolphin-mcp-pilot；案例含至少一张截图或 一段短录屏，能看到智能体请求 + dolphin-mcp-pilot 工具调用/结果。臆造、摆拍、未经验证的文字不算。
  2. 公开帖子。故事发布在至少一个公开渠道、链接可直达（不能是需登录/私密链接）；每条链接写进 frontmatter 的 channels:。帖子要有实质内容——几句话加一张图，而不是照抄本仓 README。
  3. 原创。自己的文字和截图，不抄袭，不拿 AI 生成的文字冒充第一手经历。用 AI 帮你润色文案没问题， 但发生的事情要真实。
  4. 不泄密。每张截图和配置片段都要打码 token、密码、X-DS-* 头的值、内网域名/IP、以及任何私有 业务数据，用明显的占位符替换。
  5. 案例文件要回链。你在这里提交的 README.md 必须链接到公开帖子，读者能从画廊跳到完整故事。

提交方式

  1. Fork 本仓库。

  2. 把 cases/TEMPLATE/ 复制成 cases/<你的案例 id>/（简短、唯一、kebab-case，如 nightly-etl-rerun）。

  3. 填写 README.md——补全 frontmatter 和每个小节，贴上公开链接；加一张 preview.png（或更多图） 展示智能体在操作 DolphinScheduler。

  4. 发起 Pull Request。建议标题：

    [case] <案例 id> — <一句话你做了什么>

多个案例？一个案例一个目录，放在一个 PR 里也可以。

目录结构与 frontmatter

见上文英文部分的 Directory layout 与 Frontmatter。字段含义一致，category 取上面某个创作方向， channels 至少一条公开链接。