Global Restaurant Locations Bulk Download Data: Dataset Structure and Field Coverage

Filtering happens at query time, which is exactly why the Parquet layout matters: you never need to hold all 74 million features in memory to get the 9.98 million you want. A workable walkthrough:

1. Find the latest dated release through Overture's STAC catalog instead of hard-coding a path - the August 2026 build lives under release/2026-08-19.0.
2. Point DuckDB with its httpfs and spatial extensions at s3://overturemaps-us-west-2/release/<date>/theme=places/type=place/* (or the Azure equivalent) and query object storage directly.
3. Filter on basic_category = 'restaurant' for the coarse cut, or traverse taxonomy.hierarchy for finer labels such as asian_restaurant within the roughly 2,300-category OPC taxonomy; about 280 cognitively basic labels sit at the top of that tree.
4. Apply a confidence threshold. Existence confidence runs from 0 to 1, records at or below 0.2 are already dropped before publication, and anything scoring 0 is set to operating_status = 'permanently_closed' - keep the column so your file separates live venues from dead ones.
5. COPY the surviving rows out to Parquet or GeoJSON for your own stack.

The United Kingdom is the easiest national case. The Food Standards Agency's FHRS open data exposes a keyless JSON and XML API that needs no registration, alongside per-authority bulk XML extracts refreshed daily across roughly 380 local authority feeds covering England, Wales and Northern Ireland (FHRS) and Scotland (FHIS). Volume is real even at council level - Camden alone publishes 4,260 establishment records in about 4 MB of XML - and every record carries the 0-5 rating, inspection date, geocoded address and business type for each rated food business.

New York City goes deeper per venue. NYC DOHMH Restaurant Inspection Results publishes every sustained or pending violation citation for all five boroughs - about 295,000 rows across 31 columns, exportable in six formats including CSV, GeoJSON and KMZ, with the full CSV sitting in the ~150 MB range. It refreshes daily via Socrata and requires no account, which makes it the deepest single-city official extract anywhere in this slice.

Hugging Face Datasets - Restaurant Search indexes 975 community research corpora spanning aspect-based sentiment, menu NER and task-oriented dialogue, loadable in Python or via the Hub API in Parquet, CSV or Arrow formats. Between them these cover model training well; neither substitutes for the 9.98-million-feature global footprint layer described above.

Start with the restaurants data guide for the full twelve-dataset landscape, then compare the two flagship records head-to-head in our Overture Maps Places vs NYC DOHMH comparison. If your end goal is brand-level territory mapping rather than raw points, the chain-label walkthrough pairs directly with this file, and the free restaurants datasets shortlist shows how far you can go without spending anything.