Snowflake Integration

UniverSQL is a Snowflake proxy that allows you to run SQL queries locally on Snowflake Iceberg tables and Polaris catalog, using DuckDB. You can join Snowflake data with your local datasets, without any need for a running warehouse. UniverSQL relies on Snowflake for access control and data catalog so it's complementary to your Snowflake workloads.

Getting Started

Python Installation

python3 -m pip install universql

Using virtual environments (recommended):

# Create a new virtual environment
python -m venv universql-env
 
# Activate the environment
source universql-env/bin/activate         # Mac and Linux
universql-env\Scripts\activate            # Windows
 
# Start UniverSQL
universql snowflake --account lt51601.europe-west2.gcp

Docker Installation

Pull and run the Docker image (recommended for running in background):

docker run -p 8084:8084 -v ~/.universql:/root/.universql buremba/universql snowflake --account lt51601.europe-west2.gcp

Features

Smart caching for Snowflake queries, reducing compute costs
Query local files without uploading to Snowflake for prototyping
Utilize your hardware for faster queries on small datasets
Work with your data even without internet connectivity
Use DuckDB warehouse for managed and on-premise Polaris Catalog

Cost Benefits

UniverSQL is designed to be a complementary tool that helps you utilize local hardware when using Snowflake, especially for small datasets:

X-Small warehouse costs $2/hour, while local compute is free
Only pay for egress network costs from your cloud provider
Much cheaper than running a warehouse on Snowflake
Smart caching reduces repeated data transfers

Supported SQL Operations

Operation	Behavior
`SELECT`	Executed locally using DuckDB
`CREATE ICEBERG TABLE`	`AS SELECT` executed in DuckDB; metadata synced to both catalogs
`INSERT`, `MERGE`, `DELETE`	DuckDB for temp/Iceberg tables, Snowflake for others
`CREATE TABLE`	Handled natively by Snowflake
`CREATE TEMP TABLE`	Fully managed by DuckDB
`COPY INTO`	DuckDB for temp/Iceberg tables, Snowflake for others
`SHOW`	Executed by Snowflake (metadata queries, no compute cost)

Performance

Queries that fit in memory typically run under a second
Performance comparable to Snowflake up to a certain point
Automatic routing to Snowflake for queries that can't be executed locally
Smart caching of Parquet files for better subsequent performance