Gábor Szárnyas | Data Track | Compass AI & Data Summit 2025
summary
DuckDB claims to turn laptops and phones into analytic engines: a Galaxy S24 runs a 100 GB TPC-H, a 2015 MacBook Pro handles 1 TB, and a maxed Framework laptop (8-core AMD, 128 GB) processes 10 TB. CSV ingestion with URL autodetect takes ~4s, Parquet ~1s, and local DuckDB queries drop to ~0.01s. It uses HTTP range reads for S3 pruning, a lightweight minmax index per row group, and a new DuckLake format enabling DuckDB runs on phones and DuckLake multiplayer with time travel, and it demonstrably works.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:12What is DuckDB
- 4:41Text files to databases
- 9:16DuckDB in data science
- 11:44Usability and performance
- 14:54Big data on single node
- 18:30Indexing and optimizations
- 20:21DuckDB in the cloud
- 21:23Use cases and cost savings
- 23:52Data formats and lakes
- 27:27Lakehouse architecture
- 29:15DuckLake and multiplayer setup
- 31:18DuckDB portability and demos
- 32:25Q&A
related talks
37mDuckDB: Crunching Data Anywhere, From Laptops to Servers
40mDuckDB: Harnessing in-process analytics for data science and beyond
01h 02mAnalytics for not-so-big data with DuckDB
33mData preparation and analytics on your laptop with DuckDB
27mIn Process Analytical Data Management with DuckDB
28mAnalyze data easily with duckdb - and the implications on data architectures
27mThe Duck and the DataFrame: A Data Engineer’s Journey with DuckDB
45mAnalytics for not so big data with DuckDB