Analyze Spark Job Performance with AI

Kubeflow Spark AI Toolkit

Connect AI agents and engineers to Apache Spark History Server for intelligent job analysis, performance monitoring, and investigation

This project provides two interfaces to your Spark History Server data:

📺 See it in action:

🛠️ SHS CLI (shs) - For Engineers & Scripts

A standalone Go binary. Query your Spark History Server directly from the terminal, shell scripts, or CI/CD pipelines. Also works as a skill for coding agents like Claude Code and Kiro.

Install

# Auto-detect latest version, OS, and architecture
VERSION=$(curl -s https://api.github.com/repos/kubeflow/mcp-apache-spark-history-server/releases | grep -m1 '"tag_name": "cli/' | cut -d'"' -f4 | sed 's|cli/||')
OS=$(uname -s | tr '[:upper:]' '[:lower:]')
ARCH=$(uname -m)
[ "$ARCH" = "x86_64" ] && ARCH="amd64"
[ "$ARCH" = "aarch64" ] && ARCH="arm64"

curl -sSL "https://github.com/kubeflow/mcp-apache-spark-history-server/releases/download/cli%2F${VERSION}/shs-${VERSION}-${OS}-${ARCH}.tar.gz" | tar xz
sudo mv shs /usr/local/bin/

Quick Start

# Generate a config file
shs setup config > config.yaml   # then set your Spark History Server URL

# Explore applications
shs apps
shs jobs -a APP_ID --status failed
shs stages -a APP_ID --sort duration
shs compare apps --app-a APP1 --app-b APP2

# Use as a skill with Claude Code or Kiro
shs setup skill > ~/.claude/skills/spark-history.md

CLI documentation for full usage, or check out a real-world example of Claude Code comparing two TPC-DS 3TB benchmark runs.

⚡ MCP Server - For AI Agents

An MCP (Model Context Protocol) server that exposes Spark History Server data as tools for AI agents. Agents query your Spark infrastructure using natural language - the server handles tool selection, multi-server routing, and structured data retrieval.

Use the MCP server when you want an AI agent to conduct multi-step investigations, synthesize findings across tools, or answer natural-language questions about your Spark applications.

Install

# Run directly with uvx (no install needed)
uvx --from mcp-apache-spark-history-server spark-mcp

# Or install with pip
uv tool install mcp-apache-spark-history-server
spark-mcp

The package is published to PyPI.

Coding Agent Integration

Register the server with a single command. Both examples run it over stdio via uvx.
With no config file present, the server defaults to a Spark History Server at http://localhost:18080;
point it elsewhere with a config file or SHS_SERVERS__LOCAL__URL.

Claude Code (claude mcp add):

claude mcp add --env SHS_MCP__TRANSPORT=stdio --env SHS_SERVERS__LOCAL__URL=http://localhost:18080\
  --transport stdio spark-history \
  -- uvx --from mcp-apache-spark-history-server spark-mcp

Kiro CLI (kiro-cli mcp add):

kiro-cli mcp add --name spark-history --command uvx \
  --args --from --args mcp-apache-spark-history-server --args spark-mcp \
  --env SHS_MCP__TRANSPORT=stdio --env SHS_SERVERS__LOCAL__URL=http://localhost:18080

Verify in either client with claude mcp list / kiro-cli mcp list, then ask the agent to "list the available Spark applications."

The server also ships prompts - guided, multi-step workflows you run as a command. In Claude Code: /mcp__spark-history__investigate_failure <app_id>. In Kiro CLI: /prompts investigate_failure (or @investigate_failure). See Prompts for the full list and arguments.

Passing server flags and environment

The commands above have two layers: the client's own options and the arguments/environment forwarded to spark-mcp. spark-mcp itself takes a single flag, --config / -c; everything else is set through SHS_* environment variables.

For example, to point at a remote Spark History Server with an explicit config file:

# Claude Code
claude mcp add --env SHS_MCP__TRANSPORT=stdio --transport stdio spark-history \
  -- uvx --from mcp-apache-spark-history-server spark-mcp --config ~/.config/spark-mcp/config.yaml

# Kiro CLI
kiro-cli mcp add --name spark-history --command uvx \
  --args --from --args mcp-apache-spark-history-server --args spark-mcp \
  --args --config --args ~/.config/spark-mcp/config.yaml \
  --env SHS_MCP__TRANSPORT=stdio

Configure

Basic configuration below. Create a file named config.yaml:

servers:
  local:
    default: true
    url: "http://your-spark-history-server:18080"
    auth:            # optional
      username: "user"
      password: "pass"
    include_plan_description: false   # include SQL plans by default (default: false)

mcp:
  transport: "streamable-http"   # or: stdio
  port: "18888"
  debug: false

Config file location

The server looks for its config file in the following order and uses the first one it finds:

1. The --config / -c flag (e.g. spark-mcp --config /path/to/config.yaml)
2. The SHS_MCP_CONFIG environment variable
3. ./config.yaml in the current working directory
4. ~/.config/spark-mcp/config.yaml (honors $XDG_CONFIG_HOME when set)

If none exist, the server starts with built-in defaults that can be overridden by SHS_* environment variables. When a path is given explicitly via the flag or SHS_MCP_CONFIG but the file is missing, the server fails fast instead of falling back.

Tip for MCP clients: when the server is launched by an MCP client (Claude Desktop, Kiro, etc.), the working directory is not guaranteed, so a ./config.yaml may not be found. Prefer --config / SHS_MCP_CONFIG, or place the file at ~/.config/spark-mcp/config.yaml.

Configurations can be overridden with environment variables. Nesting levels are
separated by a double underscore (__), so field names and server names may
themselves contain single underscores (e.g. SHS_SERVERS__MY_SERVER__URL maps
to servers.my_server.url).

SHS_MCP__PORT          Port for MCP server (default: 18888)
SHS_MCP__TRANSPORT     Transport mode: streamable-http or stdio
SHS_MCP__DEBUG         Enable debug mode (default: false)
SHS_MCP__ADDRESS       Bind address (default: localhost)
SHS_SERVERS__*__URL     URL for a specific server
SHS_SERVERS__*__AUTH__USERNAME
SHS_SERVERS__*__AUTH__PASSWORD
SHS_SERVERS__*__AUTH__TOKEN
SHS_SERVERS__*__VERIFY_SSL
SHS_SERVERS__*__SSL_CA_CERT    Path to a PEM CA bundle for a specific server
SHS_SERVERS__*__TIMEOUT
SHS_SERVERS__*__EMR_CLUSTER_ARN
SHS_SERVERS__*__INCLUDE_PLAN_DESCRIPTION

Multi-Server Setup

Configure multiple Spark History Servers and route queries to specific ones:

servers:
  production:
    default: true
    url: "http://prod-spark-history:18080"
    auth:
      username: "user"
      password: "pass"
  staging:
    url: "http://staging-spark-history:18080"

Agents can target a specific server per query:

"Get application <app_id> from the production server"

Custom CA bundle

For a private CA or TLS-inspecting proxy, keep verify_ssl: true and point
ssl_ca_cert to a mounted PEM bundle:

servers:
  local:
    ssl_ca_cert: "/etc/ssl/custom-ca/ca-bundle.pem"

Set the same path with SHS_SERVERS__LOCAL__SSL_CA_CERT. The CA certificates
in the bundle are added to the container's default trust roots, so include the
issuing CA and intermediates.

🏗️ Architecture

graph TB
    subgraph Clients
        A[🤖 AI Agent / LLM]
        B[👩‍💻 Engineer / Script / CI]
        C[🔧 Coding Agent - Claude Code / Kiro]
    end

    subgraph "Kubeflow Spark AI Toolkit"
        D[⚡ MCP Server]
        E[🛠️ CLI - shs]
    end

    subgraph "Spark History Servers"
        F[🔥 Production]
        G[🔥 Staging / Dev]
    end

    A -->|MCP Protocol| D
    B -->|Terminal commands| E
    C -->|shs skill file| E

    D -->|REST API| F
    D -->|REST API| G
    E -->|REST API| F
    E -->|REST API| G

Connect an AI Agent

Available Tools (19)

Example Agent Queries

• "Why is my ETL job running slower than yesterday?" → get_job_bottlenecks + list_stages + compare_job_performance
• "What caused job 42 to fail?" → list_jobs + get_stage
• "Compare today's batch with yesterday's run" → compare_job_performance + compare_job_environments
• "Find my slowest SQL queries and explain why" → list_sql_executions + get_sql_execution + compare_sql_executions

Prompts (2)

Beyond tools, the server exposes MCP prompts - reusable, multi-step investigation workflows that your agent runs as a single command. Each prompt expands into a guided sequence of tool calls (which the agent still executes one at a time), so you get a consistent, evidence-driven analysis without having to remember the right tool order.

server is optional: when omitted, the prompt searches every configured server for the application(s). Pass server="<name>" only to target a specific server or disambiguate an id that exists on more than one.

Using prompts in Claude Code

MCP prompts surface as slash commands with the format /mcp__<server>__<prompt> (note the double underscores), where <server> is the name you registered the server under (spark-history in the examples above). Type / to discover them, then pass arguments space-separated after the command:

# Investigate a failed application
/mcp__spark-history__investigate_failure spark-cc4d615a5e6b4500b8eb1e9deb48cb4e

# Target a specific configured server
/mcp__spark-history__investigate_failure spark-cc4d615a5e6b4500b8eb1e9deb48cb4e production

# Compare two applications
/mcp__spark-history__compare_applications spark-app-A spark-app-B

Using prompts in Kiro CLI

List and run prompts with the /prompts command, or type @ then Tab to autocomplete. Run a prompt by name and supply its arguments:

# Open the prompt picker (shows prompts from all MCP servers)
/prompts

# Run a prompt directly
/prompts investigate_failure

# Or use the @ shortcut
@investigate_failure

Kiro CLI prompts you for arguments interactively when the prompt declares them, so you can run /prompts investigate_failure and provide app_id (and optionally server) when asked.

📸 Screenshots

🔍 Get Spark Application

⚡ Job Performance Comparison

🚀 Kubernetes Deployment

Deploy the MCP server using Helm:

helm install spark-history-mcp ./deploy/kubernetes/helm/mcp-apache-spark-history-server/

# Production configuration
helm install spark-history-mcp ./deploy/kubernetes/helm/mcp-apache-spark-history-server/ \
  --set replicaCount=3 \
  --set autoscaling.enabled=true

See deploy/kubernetes/helm/ for full configuration options.

When deployed in Kubernetes, connect Claude Desktop via mcp-remote:

kubectl port-forward svc/mcp-apache-spark-history-server 18888:18888

📔 AWS Integration

• AWS Glue - Connect to Glue Spark History Server
• Amazon EMR - Use EMR Persistent UI for Spark analysis
• AWS Spark Troubleshooting - One-shot root cause analysis and code fix recommendations for failed Spark workloads (EMR EC2, EMR Serverless). Automatically available when AWS credentials and region are configured. See IAM setup guide for required permissions.

🔧 Development Setup

git clone https://github.com/kubeflow/mcp-apache-spark-history-server.git
cd mcp-apache-spark-history-server

# Install Task runner
brew install go-task   # macOS; see https://taskfile.dev/installation/ for others

# MCP Server
task install           # install Python dependencies
task start-spark-bg    # start Spark History Server with sample data
task start-mcp-bg      # start MCP server
task start-inspector-bg  # open MCP Inspector at http://localhost:6274
task stop-all

# CLI
cd skills/cli
task build             # build ./bin/shs
task test              # unit tests
task test-e2e          # e2e tests (starts/stops Docker SHS automatically)
task start-shs         # start SHS with CLI e2e sample data

🌍 Adopters

Using this project? Add your organization to ADOPTERS.md and help grow the community.

🤝 Contributing

See CONTRIBUTING.md for guidelines.

📄 License

Apache License 2.0 - see LICENSE.

📝 Trademark Notice

Built for use with Apache Spark™ History Server. Not affiliated with or endorsed by the Apache Software Foundation.