This guide covers serving vector tiles on-the-fly from GeoParquet files — no preprocessing or tile generation step required.

What is GeoParquet?

GeoParquet is a standardized way to store geospatial data in Apache Parquet files. It is rapidly becoming the default interchange format for large geospatial datasets:

• Overture Maps Foundation distributes its entire dataset (~200 GB) as GeoParquet
• Source Cooperative, CARTO, and Google Open Buildings all publish GeoParquet
• The data science ecosystem (pandas, DuckDB, Polars) reads Parquet natively

Quick Start

Point tileserver-rs at a GeoParquet file:

[[sources]]
id = "buildings"
type = "geoparquet"
path = "/data/overture-buildings.parquet"

Or use zero-config auto-detection — drop a .parquet file in your data directory:

tileserver-rs /data/

Configuration

Basic

[[sources]]
id = "buildings"
type = "geoparquet"
path = "/data/overture-buildings.parquet"

Full Options

[[sources]]
id = "buildings"
type = "geoparquet"
path = "/data/overture-buildings.parquet"
name = "Overture Buildings"
attribution = "© Overture Maps Foundation"
layer_name = "buildings"
geometry_column = "geometry"
minzoom = 0
maxzoom = 14

How It Works

When a client requests a tile:

1. Convert tile coordinates (z/x/y) to a WGS84 bounding box
2. Read the Parquet file with row group filtering (skip groups outside the bbox)
3. Filter individual features by bbox intersection
4. Encode matching features as MVT protobuf
5. Return the tile with Content-Type: application/x-protobuf

Client request: GET /data/buildings/14/8192/5461.pbf
    ↓
z/x/y → bbox [-74.01, 40.71, -73.99, 40.73]
    ↓
Parquet reader (row group pruning via bbox metadata)
    ↓
Feature filtering (bbox intersection)
    ↓
MVT encoding (protobuf)
    ↓
Response: 200 OK, application/x-protobuf

GeoParquet 1.1 Bbox Covering

GeoParquet 1.1 files include a bbox struct column with xmin, ymin, xmax, ymax fields per feature. tileserver-rs uses this for fast spatial filtering without deserializing geometry data.

Overture Maps data includes this column — spatial queries on 73M+ features complete in ~140ms.

Files without the bbox column still work but filter by reading all geometry data (slower on large files).

Drag-and-Drop Upload

GeoParquet files can be uploaded via the built-in drag-and-drop interface. Drop a .parquet or .geoparquet file onto the map viewer to instantly visualize it as vector tiles.

The file is uploaded to the server, registered as a GeoParquet source, and tiles are served on-the-fly.

Obtaining GeoParquet Data

Overture Maps

The largest public GeoParquet dataset. Download specific themes:

# Install the Overture CLI
pip install overturemaps

# Download buildings for a bounding box
overturemaps download \
  --bbox=-74.1,40.6,-73.8,40.9 \
  -f geoparquet \
  --type=building \
  -o nyc-buildings.parquet

DuckDB

Query and export GeoParquet from any source:

INSTALL spatial; LOAD spatial;

COPY (
  SELECT name, geometry
  FROM read_parquet('s3://overturemaps-us-west-2/release/2025-03/theme=places/**/*.parquet')
  WHERE bbox.xmin >= -74.1 AND bbox.xmax <= -73.8
    AND bbox.ymin >= 40.6 AND bbox.ymax <= 40.9
) TO 'nyc-places.parquet' (FORMAT PARQUET);

GDAL/ogr2ogr

Convert any vector format to GeoParquet:

ogr2ogr -f Parquet output.parquet input.geojson
ogr2ogr -f Parquet output.parquet input.shp

MLT Transcoding

GeoParquet sources output MVT tiles. If a client requests .mlt format, the existing MLT transcoding layer converts MVT→MLT on-the-fly — no additional configuration needed.

GET /data/buildings/14/8192/5461.mlt  →  MVT generated → MLT transcoded → response

Performance

Tips

• Use GeoParquet 1.1 with bbox covering columns for fastest filtering
• Set maxzoom to avoid generating tiles at unnecessarily high zoom levels
• Use zstd compression for best file size and read performance
• For very large files (10 GB+), consider partitioning by region