PySpark MCP Server

SQL migration assistance, AWS Glue job template generation, and Spark code optimization — as an MCP server.

Not the live-Spark pyspark-mcp package. This project is SQL → PySpark / Glue source generation, published as pyspark-tools. SemyonSinchenko/pyspark-mcp introspects a running SparkSession. A deprecated pyspark-mcp console script remains here so old configs keep working; it prints a warning, then starts this server.

What It Does

• SQL Dialect Transpilation — Convert between PostgreSQL, Oracle, Redshift, MySQL, Snowflake, and Spark SQL using SQLGlot
• PySpark DataFrame API Generation — Generate DataFrame API source text from SQL, with optimization hints
• AWS Glue templates — Job script strings, DynamicFrame conversions, Data Catalog definitions, S3 layout advice
• Batch Processing — Walk SQL files/directories and emit converted modules
• Code Review & Optimization — Pattern-based review of existing PySpark source
• Pattern Detection — Find duplicated snippets and suggest utilities

What It Doesn't Do

• Recursive CTEs → provides Spark SQL equivalent + guidance (PySpark has no native recursive CTE support)
• MERGE/PIVOT/CONNECT BY → transpiles to Spark SQL, provides DataFrame API guidance
• Perfect 1:1 DataFrame API transpilation for all SQL — complex queries get Spark SQL + recommendations
• It does not start a SparkSession, submit Glue jobs, or execute SQL
• optimize(mode="code") returns suggestions; it does not rewrite your code
• glue_s3 is a path heuristic (no AWS call, no measured speedups)
• It does not replace SemyonSinchenko/pyspark-mcp for live catalog/plans

Why this vs calling sqlglot yourself

SQLGlot already transpiles dialects. This MCP adds three things around that kernel: DataFrame-API pretty-printing with join/window/cast mappings that the conversion tests lock, Glue job boilerplate strings (bookmarks, DynamicFrames, catalog tables) so an agent can emit a file instead of assembling one, and a 14-tool FastMCP surface so an LLM picks convert / mode=sql instead of wiring sqlglot itself. If you only need sqlglot.transpile(...), use sqlglot.

Quick Start

pip install pyspark-tools
pyspark-tools

Zero-clone alternative: uvx pyspark-tools. run_server.py is a development convenience that inserts sys.path and prints startup banners. Prefer pyspark-tools in configs and production.

Example: SQL → PySpark

SELECT o.customer_id, c.name, SUM(o.amount) AS total
FROM orders o
JOIN customers c ON o.customer_id = c.id
WHERE o.status = 'paid'
GROUP BY o.customer_id, c.name

Call convert with mode=sql. Captured converter output (dialect=spark):

from pyspark.sql import SparkSession
from pyspark.sql.functions import (
    col, lit, when, count, sum, avg, min, max, countDistinct,
    coalesce, concat, datediff, date_add, to_date,
    row_number, rank, lag, lead,
)
from pyspark.sql.window import Window

# Generated from SPARK SQL
spark = SparkSession.builder.appName('SQLToPySpark').getOrCreate()

# Load table: customers
customers_df = spark.table('customers')
# Load table: orders
orders_df = spark.table('orders')

# Main query
result_df = (orders_df.alias('o')
    .join(customers_df.alias('c'), (col('o.customer_id') == col('c.id')), 'inner')
    .filter((col('o.status') == lit('paid')))
    .groupBy(col('o.customer_id'), col('c.name'))
    .select(col('o.customer_id'), col('c.name'), (sum(col('o.amount'))).alias('total')))

Exact output depends on dialect detection and fallbacks; conversion tests in tests/test_sql_conversion_fixes.py pin the important constructs. Notebook-style import * / show() is opt-in via style="notebook" on the converter.

MCP Configuration

Claude Desktop

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "pyspark": {
      "command": "pyspark-tools",
      "args": []
    }
  }
}

Hermes Agent

Add to ~/.hermes/config.yaml:

mcp:
  servers:
    pyspark:
      command: pyspark-tools
      enabled_tools: all

Docker

The image is stdio only (FastMCP over stdin/stdout). There is no HTTP server on port 8000. docker compose up is for local tests, not a health-checkable web service.

docker compose --profile test run --rm pyspark-tools-test

Tools

Fourteen routers. Each takes mode= plus a small set of fields. Old 51-tool names are not registered MCP tools (they remain as Python helpers in server.py).

convert — SQL → PySpark, batch files, PDF

convert(mode="sql", sql_query="SELECT id FROM users", dialect="postgres")
convert(mode="batch_files", file_paths=["etl/job.sql"], output_dir="out")

analyze — context, data flow, codebase, workspace

analyze(mode="sql_context", sql_content="SELECT * FROM orders o JOIN items i ON o.id = i.order_id")

optimize — suggestions only (does not rewrite)

optimize(mode="code", code="df.join(other, 'id').select('*')", optimization_level="standard")

review — code review, patterns, duplicates

review(mode="code", code="df = spark.table('t')\ndf.collect()")

glue_job — template, DynamicFrame, properties, SQL conversion

glue_job(mode="template", job_name="orders_etl", sql_query="SELECT * FROM orders")

glue_schema — detect, evolve, catalog

glue_schema(mode="detect", sample_data=[{"id": 1}], table_name="orders")

glue_s3 — path-heuristic layout advice (no AWS call)

glue_s3(mode="analyze", s3_location="s3://bucket/path", database_name="raw", table_name="orders")

glue_data — incremental, CDC, bookmarks

glue_data(mode="bookmarks", job_name="orders_etl")

refactor — patterns, utilities, pipeline

refactor(mode="utilities", code_samples=["df.filter(col('a')==1)", "df.filter(col('b')==2)"])

search — conversions, patterns, context

search(mode="conversions", query="orders", limit=10)

context — store, get, assist

context(mode="store", conversion_id="job-1", context_data={"dialect": "postgres"})

batch_status — status, cancel, active, recent

batch_status(mode="recent", limit=10)

s3_source — analyze S3 / Delta (uses host AWS credentials if boto3 is installed)

s3_source(mode="analyze", s3_path="s3://bucket/prefix")

analytics — optimization / usage stats

analytics(mode="usage", limit=20)

Security

This MCP can read local files (SQL, TXT, PDF) and, if the [aws] extra is installed, list/read S3 with the host's default AWS credentials. File tools only allow paths under the process working directory (or an explicit base_path / FileHandler(base_directory=...)). That is not a sandbox.

Run the server under a restricted OS account. Do not point it at secrets directories. Do not attach AWS credentials with write access unless you intend S3 reads via s3_source / glue_s3. Optional extras:

pip install "pyspark-tools[aws]"    # boto3 for S3/Glue catalog helpers
pip install "pyspark-tools[spark]"  # pyspark — not required at runtime; generated code only

Development

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

# Test
pytest tests/ -v --cov=pyspark_tools

# Format
black pyspark_tools tests
isort pyspark_tools tests

# Lint
flake8 pyspark_tools tests

Requires Python 3.11+ (matches the CI matrix).

Architecture

pyspark_tools/
├── server.py              # FastMCP server + helper implementations
├── consolidated_tools.py  # 14 @app.tool() routers
├── sql_converter.py       # SQLGlot-based transpilation + DataFrame API generation
├── aws_glue_integration.py # Glue job templates, DynamicFrame, Data Catalog
├── advanced_optimizer.py  # Performance analysis + optimization suggestions
├── batch_processor.py     # Concurrent file processing
├── code_reviewer.py       # PySpark code review patterns
├── duplicate_detector.py  # Code deduplication
├── data_source_analyzer.py # Data source analysis (optional boto3)
└── file_utils.py          # File I/O with allow-root checks

License

MIT — see LICENSE.

mcp-name: io.github.AnnasMazhar/pyspark-mcp