6.6 Cloud-Native Vector Formats — FlatGeobuf, GeoParquet, PMTiles

Key takeaways

• FlatGeobuf is a streamable binary format that replaces GeoJSON for large datasets.
• GeoParquet is columnar, cloud-native, and integrates with the modern analytics stack.
• PMTiles packages a tile pyramid into a single HTTP-range-readable file — no tile server required.

Introduction

Web-era GIS created demand for formats that stream from object storage without a running server. This lesson covers the three most important: FlatGeobuf, GeoParquet, and PMTiles. Together they cover vector data at every scale from web delivery to analytic warehouses.

FlatGeobuf

FlatGeobuf (.fgb) is a binary vector format built on Google's FlatBuffers serialisation library. Key properties:

• Streamable — read features one at a time without loading the whole file.
• Spatially indexed — Packed Hilbert R-tree at the front; clients can find features in a bounding box quickly.
• Single file — no sidecars.
• HTTP range reads — browsers can query remote FlatGeobuf files efficiently.
• Mutable, but append-only is the typical workflow.

Reading in Python:

1import geopandas as gpd
2[object Object]
3[object Object]
4

JavaScript:

1import { deserialize } from 'flatgeobuf/dist/geojson.js';
2const url = 'https://example.com/buildings.fgb';
3for await (const f of deserialize(url, bbox)) {
4  // f is a GeoJSON Feature
5}

When to use: medium-to-large vector data, especially when serving from object storage without a tile pipeline.

GeoParquet

Parquet is the columnar data format dominant in modern analytics (Spark, DuckDB, Snowflake, BigQuery). GeoParquet (v1.0 in 2024) adds geometry encoding via WKB inside a geometry column plus schema metadata per column.

Advantages:

• Columnar — read only the columns you need.
• Strong compression — typically 5–10× smaller than GeoJSON.
• Partitioned — split by any column (e.g., by year or country) for parallel queries.
• Integrates with DuckDB, pandas, Spark, BigQuery, Snowflake, etc.

import geopandas as gpd

1-- DuckDB
2INSTALL spatial; LOAD spatial;
3SELECT COUNT(*) FROM read_parquet('s3://bucket/parcels.parquet')
4WHERE ST_Within(geometry, ST_GeomFromText('POLYGON((...))'));

GeoParquet is now the default for data lakes holding spatial datasets. Overture Maps and Microsoft Global Building Footprints publish in GeoParquet.

PMTiles

A tile pyramid is millions of tiny files. PMTiles (Protomaps, 2022) packages an entire tile pyramid into one file that browsers can read via HTTP range requests — no tile server required.

Supports:

• Vector tiles (MVT — Mapbox Vector Tile format) for dynamic styling.
• Raster tiles (PNG/JPEG/WEBP).

Workflow:

1# Convert MBTiles to PMTiles
2pmtiles convert world.mbtiles world.pmtiles
3[object Object]
4

Client:

1import { PMTilesProtocol } from 'pmtiles';
2import maplibregl from 'maplibre-gl';
3[object Object]
4

Zero infrastructure, global scale.

FlatGeobuf vs GeoParquet vs PMTiles

When to not use them

• Small datasets (< 1 MB) — GeoJSON is still easier.
• Interchange with legacy tools — shapefile / GeoPackage still beats binary-only formats.
• Editing workflows that need row-level mutation — PostGIS or GeoPackage are better.

Cloud-native principles

The common design pattern across these formats:

1. Deterministic layout — client knows where bytes live.
2. Front-loaded index — one HTTP request (or few) finds the relevant data.
3. Range reads — transfer only the needed bytes.
4. Standard static hosting — no custom server process.

A growing ecosystem applies the same principle to other domains: STAC (metadata), Zarr (N-d arrays), COG (raster), Iceberg + Delta (tables).

Self-check exercises

Summary

• Cloud-native = front-loaded index + range reads + deterministic layout + static hosting.
• FlatGeobuf: streamable binary for features.
• GeoParquet: columnar analytics.
• PMTiles: zero-infrastructure tile delivery.
• Use alongside COG (rasters) and STAC (catalogues) for full modern pipelines.

Further reading

• flatgeobuf.org — specification and clients.
• geoparquet.org — format spec, v1.1 roadmap.
• github.com/protomaps/PMTiles — reference implementation.
• Overture Maps — exemplar GeoParquet publisher at planetary scale.