Notes from the engine room.
Notes on making data AI-ready: MCP, agent guardrails, open formats and the economics of asking your data a question.
Cloudflare R2 and Basin, joined with everything else you already have
Most data platforms start by copying everything into one warehouse, even data that never needed to move. Thunderduck sits in the middle instead: it queries data where it lives, and moves only what has to move, to the place it will be used. Today that includes Cloudflare: R2 for files and Basin for Iceberg tables, each connected with one form and joinable with S3, GCS, Azure, Glue, Postgres, MySQL, ClickHouse and BigQuery.
Read the post →DuckDB and Apache Iceberg on AWS Glue: the SQL that works and the limits
Reading Iceberg tables in AWS Glue from DuckDB takes four statements. Writing them took us three layers of debugging: Glue requires a table location, DuckDB does not send one, and Glue ignores the database's. Here is the SQL that works, the IAM it needs, and the limits we still live with.
Read the post →SQLAlchemy and Superset for DuckDB at scale: pandas, Polars and PyTorch
thunderduck-sqlalchemy is on PyPI: a SQLAlchemy dialect and DBAPI that turns your lakehouse into a data source anything Python already knows how to read. pip install, one URL, and pandas, Polars and PyTorch are reading governed tables with no extract in between — and the same URL points Apache Superset at the engine for ad-hoc BI.
Read the post →Zero ETL first — then ETL, when you actually need it
Federated queries delete most of the pipelines you were about to build: no copies, no sync lag, one SQL statement across your lake and your databases. But some workloads genuinely want a materialized table — so Thunderduck runs CREATE TABLE AS, incremental loads and full dbt Medallion modeling on the same engine, against the same catalogs.
Read the post →Hello world: the birth of Thunderduck
Why Thunderduck exists: making your data AI-ready should take an afternoon, not two quarters and a five-figure monthly bill. Our mission, our vision, and the two questions every company should be asking about its data platform.
Read the post →Your data, AI-ready by lunch.
And at a fraction of warehouse cost — that part still matters.