Coming from tippecanoe¶

tippecanoe is tylertoo's tiling reference. tylertoo applies the same concepts to stored overview levels in world space before encoding tiles. context/ARCHITECTURE.md describes the deliberate differences. benchmarks/e2e/RESULTS.md compares speed, size, and memory use, with methods and caveats.

Like tippecanoe, tylertoo thins points by default and keeps lines and polygons. This preserves polygon coverage at coarse and middle zooms. Polygons too small for a zoom become placeholder squares. For oversized tiles, tylertoo merges dense polygons into their neighbors, as --coalesce-densest-as-needed does. To adjust this behavior, see complete tiles or small tiles.

How much faster — the skeptical reader's table¶

Measured end-to-end (GeoParquet in, z0–z14 PMTiles out) against tippecanoe v2.79.0 built from a verified pinned commit, on the corpus datasets with tippecanoe running its recorded flags for each (--drop-densest-as-needed, --simplification=10, …) — both tools in a production configuration, not a rigged one. The "thinning off" column re-runs tylertoo with --verbatim --simplify-factor 1.0 so it cannot win by emitting fewer features. Apple M3 Pro, median of 2 warm runs; one recorded run on a non-idle machine, so treat ratios as approximate.

The honest footnotes, which are the point of this table:

• † The point row is not a parity measurement at all. tippecanoe's recorded flags had it building z15–z16 — 9,639 of its 10,486 tiles and 68% of its output bytes — while tylertoo stopped at the requested z14. Both of that row's ratios include that extra work.
• The point row's "thinning off" 2.4× is the weakest number this harness has produced, and it carries a second mismatch on top of †: --verbatim makes tylertoo carry all 598k points at every zoom (1.39 s → 12.21 s) at 2110 MB peak RSS, 4.07× tippecanoe's. Both tools drop points by design at production settings, so neither point ratio describes how anyone would actually run them. Read the line and polygon rows for parity numbers.
• tylertoo's peak memory is higher on three of four datasets — up to 2.5× on lines. The speed is parallelism (6–7 cores busy vs tippecanoe's ~2) plus the native GeoParquet read, paid for in working set.
• Archive sizes are smaller or equal except Portland buildings (1.11× at defaults, 1.20× thinning off) — and except points, where the published 0.59× compares tylertoo's z0–14 archive against tippecanoe's z0–16. At z0–14 on both sides tylertoo's points archive is 1.87× bigger.
• Converting GeoParquet for tippecanoe is only 1–2% of its end-to-end wall at this scale — the gap is tiler vs tiler, not a format-conversion tax.
• On a small contiguous admin coverage (the 17k-polygon fixture set), the parity-matched ratio is ≈7.5×, not 16–20×.

Every number above is from a committed record (benchmarks/e2e/corpus_results.json, results.json) with full provenance, per-zoom tile counts, and the asymmetry list in benchmarks/e2e/RESULTS.md. Reproduce with ./benchmarks/e2e/run_corpus.sh.

Flag mapping¶

What only tylertoo does¶

• Reads GeoParquet directly. It decodes required columns, prunes row groups with --bbox and --filter, and reads remote objects by byte range. See remote reads.
• Stores world-space overview levels in GeoParquet. Query them with DuckDB, export them multiple times, or validate them against the geo:overviews spec.
• Splits builds across machines and merges shards without re-encoding tiles. See scaling.

What only tippecanoe does¶

• Reads GeoJSON, line-delimited GeoJSON, FlatGeobuf, point CSV, and GeoJSON on standard input. tylertoo reads GeoParquet in EPSG:4326 or EPSG:3857. Convert other formats with gpio.

• Joins attributes onto finished tiles with tile-join. In tylertoo, join before tiling so the new columns work with --filter, --include-property, and --feature-id. Restore Hilbert order and row-group layout after the join:

  duckdb -c "LOAD spatial; COPY (
    SELECT p.*, c.population
    FROM 'parcels.parquet' p
    JOIN 'census.csv' c USING (geoid)
  ) TO 'joined.parquet' (FORMAT parquet)"
  gpio sort hilbert joined.parquet parcels-joined.parquet \
    --row-group-size-mb 128
  

• Writes MBTiles or a tile directory. tylertoo writes one PMTiles archive; convert it with pmtiles-convert from the Python pmtiles package. The output path selects the format. A directory contains gzip-compressed {z}/{x}/{y}.mvt files and metadata.json:

  uvx --from pmtiles pmtiles-convert tiles.pmtiles tiles.mbtiles
  uvx --from pmtiles pmtiles-convert tiles.pmtiles tiles/