How this works
- Parquet on plain file storage. Each dataset is a single columnar Parquet file (ZSTD-compressed, sorted on its main date column, statistics in every row group). In production these files would sit on object storage like Cloudflare R2 or S3; here they ship with the page.
- Range requests, not downloads. DuckDB reads the file footer first, then fetches byte ranges for the column chunks a query touches. A count(*) is answered from footer metadata alone — kilobytes, not megabytes — and a date filter skips whole row groups without reading them. The “What it read” view above shows this happening, from the run itself.
- The engine is WebAssembly. DuckDB-WASM — the official DuckDB project, also behind MotherDuck's browser UI and Evidence.dev sites — runs the full analytical engine in a Web Worker in this tab. Your SQL never leaves the browser.
What it can't do
- No shared state. Every visitor gets their own engine. There is no multi-user warehouse here — no concurrency, no shared catalog, no accounts.
- Uploads live for the session. Files you add are held in this tab's memory and are gone when it closes — by design. Re-adding a file under the same name is the refresh pattern, and saved queries that reference a missing upload say so and ask for the file again.
- Bounded by the tab. Working sets have to fit in browser memory — keep them in the low gigabytes and lean on predicate pushdown.
- Two honest qualifications. Importing from a link contacts that link's host from your browser. A ClickHouse connection holds its credentials in this tab only — use it for your own database, ideally with a read-only user.
Data: NYC TLC yellow taxi trips, Jan–Feb 2024, trimmed from the source's 19 columns and sorted by pickup time. Engine: DuckDB-WASM.