Synthetic star-schema data generator for dbt + DuckDB. Generates 7 Parquet files (~500 MB) with 50K–5M rows each.

# Clone and setup
git clone https://github.com/alwyndsouza/dbt_synthetic_data.git
cd dbt_synthetic_data

# Full setup: uv environment, dbt packages, and synthetic data
make setup

This generates:

data/raw/
├── dim_users.parquet           (50K rows)
├── dim_products.parquet        (200 rows)
├── dim_locations.parquet       (500 rows)
├── dim_devices.parquet         (100 rows)
├── fact_transactions.parquet   (1M rows)
├── fact_sessions.parquet       (500K rows)
└── fact_events.parquet         (5M rows)

make env-setup    # Setup uv environment and dbt packages
make data         # Generate synthetic Parquet files
make install      # Create uv environment and install dependencies
make dbt-deps     # Install dbt packages
make setup        # Full setup (env + packages + data)
make clean        # Remove data and dbt artifacts
make help         # Show all commands

• generate_data.py - Data generation script
• dbt_project.yml - dbt project config
• profiles.yml - DuckDB connection settings
• packages.yml - dbt package dependencies
• models/staging/sources.yml - External Parquet table definitions

Query Parquet files directly in your models:

select * from read_parquet('data/raw/dim_users.parquet')

Sources are documented in models/staging/sources.yml for dbt's source freshness checks.

• Referential integrity — all FK values are validated against their PK sets before saving.
• Temporal consistency — signup_date < transaction_date; event_timestamp falls within its session window.
• Business logic — premium users have 3× transaction volume; enterprise users have 10% higher values; failed transactions have $0 tax; mobile devices generate 60% of events.
• Realistic distributions — weekday transaction peaks, 40% weekend event drop, Q4 seasonal spike.

import duckdb

con = duckdb.connect()

# Inspect a dimension table
con.execute("SELECT * FROM read_parquet('data/raw/dim_users.parquet') LIMIT 5").df()

# Check referential integrity
con.execute("""
    SELECT COUNT(*) AS orphan_transactions
    FROM read_parquet('data/raw/fact_transactions.parquet') t
    LEFT JOIN read_parquet('data/raw/dim_users.parquet') u USING (user_id)
    WHERE u.user_id IS NULL
""").fetchone()