smallpond

Project Overview

This repository is a fork of https://github.com/deepseek-ai/smallpond, enhanced with additional support for duckdb engine to aceess 3FS using hf3fs_usrbio.

Using hf3fs_usrbio with DuckDB Engine

Ensure the 3FS storage cluster is mounted at /3fs/ , so duckdb engine use hf3fs_usrbio to acess files in 3FS, and arrow engine use standard POSIX to acess files in 3FS.

import smallpond

sp = smallpond.init(data_root="/3fs/duckdb/data_root/")
df = sp.read_parquet("/3fs/duckdb/prices.parquet")
df = df.repartition(3, hash_by="ticker")
df = sp.partial_sql("SELECT ticker, min(price), max(price) FROM {0} GROUP BY ticker", df)
print(df.to_pandas())
df.write_parquet("/3fs/duckdb/output/")

Features

Smallpond is a lightweight data processing framework built on DuckDB and 3FS.

• 🚀 High-performance data processing powered by DuckDB
• 🌍 Scalable to handle PB-scale datasets
• 🛠️ Easy operations with no long-running services

Installation

Python 3.8 to 3.12 is supported.

pip install smallpond

Quick Start

# Download example data
wget https://duckdb.org/data/prices.parquet

import smallpond

# Initialize session
sp = smallpond.init()

# Load data
df = sp.read_parquet("prices.parquet")

# Process data
df = df.repartition(3, hash_by="ticker")
df = sp.partial_sql("SELECT ticker, min(price), max(price) FROM {0} GROUP BY ticker", df)

# Save results
df.write_parquet("output/")
# Show results
print(df.to_pandas())

Documentation

For detailed guides and API reference:

• Getting Started
• API Reference

Performance

We evaluated smallpond using the GraySort benchmark (script) on a cluster comprising 50 compute nodes and 25 storage nodes running 3FS. The benchmark sorted 110.5TiB of data in 30 minutes and 14 seconds, achieving an average throughput of 3.66TiB/min.

Details can be found in 3FS - Gray Sort.

Development

pip install .[dev]

# run unit tests
pytest -v tests/test*.py

# build documentation
pip install .[docs]
cd docs
make html
python -m http.server --directory build/html

License

This project is licensed under the MIT License.