Starlet: An End-to-End Python Library for Scalable Geospatial Tiling and Vector-Tile Serving

Authors: Tarlan Bahadori, Rohan Bennur, Shaolin Xie, Ibrahim Sabek, Ahmed Eldawy

Accepted at ACM SIGSPATIAL 2026 (Riverside, CA, Nov 3–6) — Short Paper.

Starlet is a pip-installable Python library for end-to-end geospatial tiling and vector-tile serving over GeoParquet. It ingests GeoParquet or GeoJSON, partitions the data into balanced spatial shards, builds a Web Mercator density grid for fast tile-occupancy tests, and generates Mapbox Vector Tiles either eagerly into an on-disk pyramid or lazily on first request.

Why Starlet

Web-scale geospatial visualization normally takes three stages—transform to a tile-friendly layout, construct or select a pyramid, and serve tiles to a renderer—and existing systems address these separately. Few offer an end-to-end path that is native to modern columnar spatial data and easy to embed in a Python workflow. Starlet packages ingestion, tile generation, serving, and prompt-based styling into a single library that runs on one machine.

Key Capabilities

• R*-Grove-inspired partitioning with a round-buffered streaming writer that bounds open files and per-round memory during ingestion
• 4096×4096 prefix-sum density grid answering rectangular range counts in O(1), so empty and low-density tiles are pruned before geometry decoding
• Eager or lazy MVT generation, with a serving layer combining memory caching, disk-backed tiles, and on-demand generation
• PMTiles export cutting on-disk storage by 47–64% versus loose .mvt files
• LLM-assisted style reuse and prompt sharing over a repository of annotated visualization templates

Evaluation Highlights

• Partitions and indexes a 17.6 GB OpenStreetMap extract — 43M features, one billion vertices — in 38.5 minutes, peak RSS below 9.6 GB, on commodity laptop hardware
• Cached tile serving stays interactive under load: median latency 4.7 ms at 20 concurrent clients, p95 below 5 ms
• Runtime grows linearly with feature count (397 s → 6,397 s as input grows 17×), while peak memory grows sub-linearly

Artifacts:

• 📄 Paper: /files/starlet_sigspatial26_short.pdf
• 📦 Package: pypi.org/project/starlet — pip install starlet
• 💻 Code: github.com/ucr-bdlab/starlet