PMTiles and COG Killed My Tile Server

What previously required PostgreSQL + PostGIS, GeoServer, a load balancer, and Redis caching now runs from a single file on S3 at $0.02/GB/month with zero compute costs during idle periods. PMTiles packages all vector tiles into one Hilbert-curve-indexed file. COG structures raster imagery for HTTP Range Requests. Neither requires a running server. Tippecanoe achieves 92% size reduction from raw GeoJSON, translating to 6x faster loads on mobile networks.

Tile Servers: High Ops Cost, Poor Idle Economics

Traditional tile servers accept requests, query a database, generate tiles dynamically, and serve them. This requires managing database connections, caching strategies, horizontal scaling during traffic spikes, and paying for compute during idle periods. For government agencies and research institutions with limited DevOps capacity, the overhead often means geospatial data stays locked in desktop GIS.

COG: Sub-Second Previews of Multi-Gigabyte Satellite Scenes

A Cloud Optimized GeoTIFF is a standard GeoTIFF reorganized with internal tiling (256x256 or 512x512 blocks) and pre-computed overviews. This structure enables HTTP Range Requests -- clients fetch only the bytes they need. Previewing a multi-gigabyte satellite scene takes under a second. No tile server. Just an S3 bucket and CloudFront.

COG Technical Properties

PMTiles: One File Replaces Your Entire Tile Server

PMTiles solves the same problem for vector data. Developed by Brandon Liu at Protomaps, it packages all vector tiles into a single file served from static storage -- S3, R2, Azure Blob, or a CDN. No database, no tile generation, no caching layer.

Hilbert Curve ordering ensures spatially adjacent tiles are physically adjacent in the file, enabling efficient range requests for any viewport. A world-scale basemap PMTiles file runs 50-100GB, but requesting tiles for a city view fetches just kilobytes.

GeoParquet: Columnar Analytics with DuckDB, Spark, and BigQuery

COG and PMTiles optimize for visualization. GeoParquet targets analytics -- spatial joins, aggregations, ML feature engineering. It inherits Apache Parquet's columnar compression, predicate pushdown, and integration with DuckDB, Spark, BigQuery, and Snowflake. Geometry is stored as WKB, compatible with virtually every GIS tool.

Which Format for Which Job

• COG — Raster visualization and processing (satellite imagery, DEMs, aerial photos)
• PMTiles — Vector visualization (basemaps, reference layers, thematic maps)
• GeoParquet — Vector analytics (spatial joins, aggregations, ML feature engineering)
• FlatGeobuf — Simple vector streaming with bbox filtering
• Zarr — Multi-dimensional arrays (climate data, time series rasters)

Tippecanoe: 92% Smaller Than Raw GeoJSON

Mapbox's Tippecanoe is the gold standard for PMTiles generation. It strips features at low zoom levels that would render as single pixels and simplifies geometries to match per-scale visual fidelity. The result: 92% size reduction compared to raw GeoJSON, translating to 6x faster loads on mobile networks.

# Generate PMTiles from GeoJSON
tippecanoe -o output.pmtiles \
  --minimum-zoom=0 \
  --maximum-zoom=14 \
  --drop-densest-as-needed \
  --extend-zooms-if-still-dropping \
  input.geojson

Production Stack: Five Components, Zero Servers

The Serverless Geospatial Stack

Cost comparison: multiple EC2 instances + RDS + Redis cache versus commodity storage at $0.02/GB/month with zero compute costs during idle periods.

The One Caveat: Transactional Editing Still Needs a Database

These formats optimize for visualization and analysis, not transactional editing. Frequent spatial data updates still require a database. But even then, PostGIS can export to cloud-native formats on a schedule for read-heavy workloads. The migration path for organizations still running tile servers: convert to COG/PMTiles, upload to object storage, update client code, decommission servers.