PMTiles Pipeline

Reference for building a single global PMTiles archive on one fat node, distilled from a real global change-layer build (~6T-px raster input, 176M polygons, 28 GB PMTiles output). The pipeline is:

raster.zarr  →  geoparquet shards  →  FlatGeobuf batches  →  sharded tippecanoe  →  tile-join  →  pmtiles
   (1)              (2)                    (3)                     (4)                  (5)

Each stage's "why" matters more than the syntax. Use this skill whenever a step in this pipeline is being built, debugged, or optimized.

Stage 1 — Raster → polygons (as parquet shards)

• Use contourrs.shapes_arrow to vectorize a classified uint8 raster directly to Arrow (skip GeoJSON in/out).
• Pre-filter with scipy.ndimage.binary_opening(struct=NxN) per class to wipe salt-and-pepper noise. N=3 is a sane default at 10m resolution.
• Write GeoParquet via geopandas with compression="zstd". One parquet per super-tile (e.g., 8192² px) → tens of thousands of small shards. That granularity matters for downstream resume + parallelism.
• Track fields: class (uint8), area_deg2 (float64 — useful for downstream filtering), geometry.
• Connectivity 8 (diagonals count) and min-pixels=25 are fine starting defaults; higher min-pixels = aggressive smoothing.