Skip to content

Demo: Brazil 2025 field predictions, filtered from a 630 GiB collection

A tiling run that starts from the global Fields of The World predictions collection on Source Cooperative — 629.6 GiB of GeoParquet, 8.2 billion rows — and ends with a z0–14 PMTiles pyramid of Brazil's 2025 growing-season field predictions: --filter selects one class and one vintage, --bbox selects the country, and --representation "0-7:point" renders dots zoomed out, polygons zoomed in, all in one archive.

There is no established tool that produces this archive. Tippecanoe — the standard vector tiler — does not read GeoParquet, let alone filter a remote collection while tiling it. tylertoo reads the cloud-native source, carves the slice, and writes the pyramid in one workflow.

Explore the tiles

43.9M field predictions, rendered live from the PMTiles archive on CARTO Dark Matter. Zoomed out you see the centroid band (z0–7); from z8 the actual field polygons take over. Pan and zoom, or jump to one of Brazil's agricultural heartlands with the buttons.

Open the map full-screen →

Tiles are served from Source Cooperative with HTTP Range + CORS (nlebovits/gpq-tiles-demo).

The measured run

Full remote → PMTiles round trip on a 16-core machine, zoom z0–14. The input is the 52 part files of the global collection whose footers intersect Brazil (40.7 GiB read over HTTPS); the other 948 files (589 GiB) were never touched.

stage wall peak RSS output
convert (incl. 40.7 GiB remote read) 1h 11m 56s 9.6 GiB 14.1 GB overview GeoParquet (15 levels)
export 11 m 44 s 1.54 GiB 4.5 GiB PMTiles, 1,647,927 tiles
total 1 h 23 m 40 s — z0–14, 116,504,741 tile-features, 0 oversized 
# 1. Overview straight from the remote collection slice:
#    2025 vintage, fields only, Brazil bbox, centroids at z0-7.
tylertoo overview --files-from brazil-2025-manifest.txt \
  --bbox="-74.1,-34.0,-34.7,5.4" \
  --filter "label = 'field' AND time >= '2025-01-01'" \
  --min-zoom 0 --max-zoom 14 \
  --representation "0-7:point" \
  brazil-2025-fields-ov-z14.parquet

# 2. Export the PMTiles archive.
tylertoo export-pmtiles brazil-2025-fields-ov-z14.parquet \
  brazil-2025-fields.pmtiles

What the run shows

  1. The filter is the feature. No curated "Brazil 2025 fields" extract exists — the slice lives interleaved with two other prediction classes and the 2024 vintage inside 1,000 Spark part files. The predicate label = 'field' AND time >= '2025-01-01' plus the bbox kept 43.9M of the 426M rows scanned (10.3%), evaluated during the tiling read — no DuckDB pre-pass, no intermediate file.
  2. One archive serves dots and polygons. The z0–7 band stores one representative point per surviving feature (a point is always visible, so even z0 renders — the previous polygon-only demo's z0 was empty); z8–14 store the polygons. A two-line style split (circle + fill) renders both.
  3. Cloud-native at collection scale. The run fetched 6.9% of the collection's bytes — the 52 files that could contain Brazil — once each (~1.0×, spilled locally so later passes never re-hit the network).
  4. Bounded memory at every stage. Convert auto-selected spill mode (the in-RAM estimate for its 109M output rows was ~315 GiB) and peaked at 9.6 GiB; export streamed all 15 levels at a peak of 1.54 GiB.

Full methodology, per-zoom breakdown, and the upstream findings the run surfaced (INT96 timestamps, "crs": null, missing covering metadata) are in the demo directory on GitHub.

Clipping/simplification differ from tippecanoe by design

This demonstrates the native GeoParquet pipeline and its output, not byte-identical tiling. See Architecture for the documented divergences.