DuckDB

DuckDB is “SQLite for Analytics”. It is an in-process SQL OLAP database. It runs inside your application process and is blazing fast for analytical queries on local files (Parquet, CSV, JSON).

When to Use

• Local Analytics: Analyze millions of rows on your laptop in seconds.
• Data Engineering: Process data in Python/R pipelines (replacement for Pandas).
• Serverless Data Lake: Query S3 parquet files directly via Lambda without a running warehouse.

Quick Start (Python)

import duckdb

# Query local CSV directly
duckdb.sql("SELECT avg(price) FROM 'sales.csv' WHERE region='US'").show()

# Connect to S3
duckdb.sql("INSTALL httpfs; LOAD httpfs;")
duckdb.sql("SELECT count(*) FROM 's3://my-bucket/data.parquet'")

Core Concepts

Vectorized Execution

Standard DBs process row-by-row. DuckDB processes batches of columns (Vectors), utilizing modern CPU SIMD instructions.

Universal Format Reader

Can query CSV, JSON, Parquet, Arrow, SQLite, and Postgres tables as if they were local tables.

Zero Dependencies

Single binary/library.

Best Practices (2025)

Do:

• Use Parquet: It is the native language of analytics. DuckDB + Parquet is incredible.
• Replace Pandas: For datasets larger than RAM, DuckDB works (Disk spilling) where Pandas crashes.
• Use explicitly typed SQL: DuckDB’s SQL dialect is very friendly and standard (Postgres-compatible).

Don’t:

• Don’t use for Multi-User OLTP: It handles concurrency poorly (single writer). Use Postgres for that. Use DuckDB for analysis.

References

• DuckDB Documentation