tao.chenandClaude f374ebfdb4 uploads,tools,config,main: 将 upload_file 返回路径改为服务器端绝对路径并透传给 spark_submit
之前 upload_file 把文件写到 DataDir/uploads/<filename>,返回相对路径
uploads/hello.txt。MCP 服务器与 agent 不在同一台机器,agent 无法构造服务
器本地路径,因此 spark_submit 无法定位脚本。

改为由新的 internal/uploads 包 mint 一个 32 字符十六进制 file_id,文件落盘为
DataDir/uploads/<file_id>,元数据写入 <file_id>.meta.json(原始文件名只存在
sidecar 里)。upload_file 返回 {file_id, path, name, size, sha256},其中 path
是绝对路径。spark_submit 的 description 明确要求 LLM 直接把 upload_file 返回
的 path 放进 args,不要自己构造路径。

为什么只在描述里约束而不在代码里拒绝非 mint 的绝对路径:集群本地已有路径
(如 /opt/spark/examples/pi.py)是合法的 spark-submit 参数,代码不能替
LLM 拒绝。

测试锁定:
- internal/uploads: Save 往返、AbsPath/Validate 非法路径、Sweep 过期/未过期/
  孤立 sidecar
- internal/mcp/tools: upload_file 新响应字段、spark_submit 透传 mint 路径与
  集群本地路径

刻意未做:S3/HDFS 上传、给 spark_submit 新增 file_id 参数、在代码层面拒绝
集群本地绝对路径。

破坏性变更:upload_file 响应从相对 path 改为绝对 path,并新增 file_id/name/
sha256 字段。该工具尚无外部调用者。

Co-Authored-By: Claude <noreply@anthropic.com>
2026-07-13 19:13:14 +08:00

spark-mcp-go

MCP (Model Context Protocol) server for Apache Spark on YARN. Lets an LLM agent discover Spark/YARN endpoints, submit jobs, fetch logs, and analyze them via 11 typed Tools. Streamable HTTP transport, SQLite-backed configuration, and log/slog structured logging with per-Tool call files.

What it does

  • Exposes 11 MCP Tools over Streamable HTTP at /mcp
  • Admin API at /admin/* for cluster and audit configuration
  • HTTP Basic and YARN SimpleAuth for Spark/YARN endpoints
  • SSRF protection with cluster URL allowlist plus DNS rebinding guard
  • Per-Tool independent audit log (file mode 0600)
  • spark-submit via local exec.Command (no shell, no injection)

Quick start (3 steps)

  1. Copy and edit the environment file:

    cp .env.example .env
    # Edit ADMIN_TOKENS and AGENT_TOKEN
    
  2. Build and start the server:

    go build ./...
    ./spark-mcp-go
    # Or use the helper:
    # ./scripts/dev.sh
    
  3. Initialize an MCP session, then call list_clusters to discover endpoints.

11 Tools

Tool Type Purpose
list_clusters discovery Return all active clusters (LLM entry point)
spark_submit exec Local spark-submit process, extracts app_id
list_applications RM GET /ws/v1/cluster/apps[?state=&user=]
get_application_status RM GET /ws/v1/cluster/apps/{id}
get_application_logs RM Fallback chain: amContainerLogs -> aggregated-logs -> logs
kill_application RM PUT state KILLED
fetch_spark_metrics SHS Executor metrics + summary mode triggers analyzer
fetch_cluster_env RM Aggregate /cluster/info and /cluster/metrics
analyze_spark_log analyzer RM logs + heuristic rules + LLM-ready prompt
fetch_url HTTP Generic primitive, uses cluster allowlist and auth
upload_file FS Write to ./data/uploads/, reference from spark_submit

Environment variables

Variable Default Required Description
LISTEN_ADDR :8080 no HTTP listen address
DATA_DIR ./data no SQLite, uploads, and log root
ADMIN_TOKENS yes Comma-separated tokens for /admin/*
AGENT_TOKEN yes Token for /mcp requests
HTTP_CLIENT_TIMEOUT 30s no Timeout for fetch_url/RM Tool HTTP calls
MAX_RESPONSE_BYTES 1048576 (1 MiB) no HTTP response truncation limit
SPARK_SUBMIT_TIMEOUT 60s no spark-submit process timeout
LOG_DIR ./data/logs no Log root directory
LOG_LEVEL info no debug, info, warn, or error
LOG_FORMAT text no text for terminal, json for log files
ANALYZER_DATA_SKEW_RATIO 3.0 no Data-skew rule threshold (max/min ratio)
ANALYZER_GC_PRESSURE_RATIO 0.1 no GC-pressure rule threshold (GC/CPU ratio)
ANALYZER_BOTTLENECK_SHUFFLE_GB 50 no Bottleneck rule threshold (shuffle GB)

Admin API

All endpoints require Authorization: Bearer <admin_token>.

Method Path Description
GET /admin/clusters List all clusters
POST /admin/clusters Create a cluster
GET /admin/clusters/:id Get one cluster
PUT /admin/clusters/:id Update a cluster
DELETE /admin/clusters/:id Delete a cluster
GET /admin/audit?limit=100 Query audit log

Create a cluster:

curl -sS -X POST \\
  -H "Authorization: Bearer ${ADMIN_TOKEN}" \\
  -H "Content-Type: application/json" \\
  -d '{
    "id": "prod",
    "name": "prod",
    "rm_url": "http://rm.example.com:8088",
    "shs_url": "http://shs.example.com:18080",
    "spark_submit_execute_bin": "/opt/spark/bin/spark-submit",
    "is_active": true,
    "auth_type": "simple",
    "auth_username": "yarn",
    "rate_limit_per_min": 10,
    "url_allowlist": ["rm.example.com:8088", "shs.example.com:18080"]
  }' \\
  "http://127.0.0.1:${LISTEN_ADDR:-:8080}/admin/clusters"

Note: auth_password is not accepted via JSON (json:"-"). Set it through a dedicated password endpoint or seed the database directly.

Security model

Three lines of defense:

  1. Token auth - ADMIN_TOKENS protects /admin/*; AGENT_TOKEN protects /mcp.
  2. SSRF guard - Every outbound URL must match the cluster's url_allowlist, with private-IP CIDR blacklist (incl. 169.254.0.0/16 AWS/GCP metadata).
  3. Path allowlist - upload_file writes only to ./data/uploads/, rejects absolute paths, .., and non-[a-zA-Z0-9._-] filenames.

Additional guarantees:

  • spark_submit runs the cluster's binary via exec.Command(name, args...) — slice form, never sh -c. No shell metacharacter interpretation.
  • auth_password is tagged json:"-"; never serialized in responses, never accepted from JSON request bodies.
  • Cross-host redirects (RM → NM 307) preserve the Authorization header through DoWithRedirect only when the destination host is in url_allowlist.
  • Per-Tool log files are created with mode 0600 and live under ./data/logs/tools/.

Development

go build ./...
go test ./...
go vet ./...

More docs

S
Description
No description provided
Readme
268 KiB
2026-07-15 16:37:20 +08:00
Languages
Go 86.5%
HTML 12.6%
Shell 0.9%