Changelog¶

Notable changes to this project are recorded here.

The format is based on Keep a Changelog, and releases follow Semantic Versioning.

Unreleased¶

Changed¶

Rust callers pass a TileCoord to StreamingPmtilesWriter::add_tile_precompressed instead of separate zoom, x and y arguments. Compression::from_str is now Compression::parse; str::parse::<Compression>() is also available.
Remote staging reports each operation's live concurrency, including later operations in the same process (#779).

v0.8.0 (2026-10-06)¶

Sharded builds, much faster conversion and export, and a long list of robustness fixes. Some commands now refuse to overwrite existing files; see Changed before upgrading.

During development, main carried versions 0.8.0 to 0.11.0. None of those were released. This is the first release after v0.7.1.

Added¶

Sharded builds: tiles --shard I/N builds one disjoint slice of a large job, --tile-range limits a build to part of the tile space, and tylertoo merge joins the shards into one archive. The merged result matches a single-machine build (#498, #510, #537, #541, #555).
Saved convert plans: --save-plan writes the expensive first pass to a file, --plan reuses it, and tiles --plan-only writes the plan without exporting tiles (#505, #521, #560, #574, #600).
--max-zoom auto picks the finest zoom from the data, similar to tippecanoe's -zg. In Python, pass max_zoom="auto" (#444, #573).
--feature-id <COLUMN> sets stable MVT feature ids from a column, so MapLibre setFeatureState works across tiles and zooms (#443, #568).
tylertoo stats reports tile sizes per zoom (count, total, mean, p50, p99, max), with --largest N and --json (#551, #557).
pyramid --feature-order, and --band now accepts remote inputs and paths that contain colons (#374, #491, #598).
Nested properties: struct, list and map columns reach the tiles as JSON strings instead of being dropped. Columns that are still skipped produce a warning (#434, #599).
--bbox on GeoParquet 2.0 files skips row groups using the file's native geometry statistics (#497, #500).
Profiling: TYLERTOO_PROFILE_JSON writes per-phase timers and memory use, including peak memory within each phase (#502, #550, #571, #627).

Changed¶

Existing outputs are no longer overwritten by default. tiles, the bare tylertoo IN OUT form, overview, export-pmtiles and decode refuse to write over an existing file unless you pass -f/--force. A directory at the output path is always refused (#427, #552, #595).
Zoom limit: zooms above 30 are rejected up front. Before, they were accepted and the run hung (#371, #488).
Out-of-range input fails loudly: when every feature falls outside the valid coordinate range (for example, meters labelled as degrees), the conversion fails instead of writing an empty archive. Partial drops are counted and reported, and warnings from the core library now reach the CLI (#429, #493, #553, #563).
Clustered PMTiles: tiles are written in tile-id order, and the header min and max zoom match the tiles actually stored, so go-pmtiles verify accepts the archives (#501, #506, #516, #529, #554).
Pyramid bands are checked against the tiles they really contain, not their declared zoom range (#495, #503, #514, #527).
Rust API: tylertoo-core has breaking changes. The level-assignment and cluster functions take &FeatureTable instead of &[AssignFeature]; see crates/core/api/tylertoo-core.txt for the full surface (#543, #630).

Fixed¶

Features that cross the antimeridian keep the part past ±180°, as in tippecanoe (#342, #642).
At zoom 15 and above, the top rows of tiles were lost because latitude was clamped at ±85.05 instead of the exact Web Mercator limit (#416, #484).
On the musl Linux binary, large runs no longer abort from memory fragmentation (#480, #483).
--profile auto and --partition-wave auto respect container memory limits (cgroup v1 and v2, as used by Docker, Kubernetes and Slurm) (#481, #485).
Killed runs no longer leave a truncated overview or decode output in place of the previous file (#427, #595).
GeometryCollections are encoded as one feature per geometry type instead of being dropped (#431, #592).
Hostile or corrupt input is rejected with an error instead of a crash or wrong output: malformed WKB, truncated or hostile remote Parquet footers, hostile PMTiles archives, out-of-tile coordinates, NaN values, and invalid --extent or --tile-buffer values (#399, #406, #417, #428, #430, #432, #433, #623, #632).
Simplifying very long lines no longer overflows the stack (#575, #578).
Output is byte-identical across thread counts (#423, #487, #609, #610, #611).
The first pass fails early with a clear message when its feature table cannot fit in memory, and pass-1 line memory stays within its limit (#449, #543, #549, #569).
Writing more than 32,767 row groups now works (#507, #508, #509).

Performance¶

Export is about 1.75× faster (#535, #559), and checkpoints cost the same no matter how large the archive is (#459, #528).
The first pass and the level assignment run in parallel (#460, #504, #534, #564).
Pass-2 reads run in parallel, and spilling to disk no longer blocks (#494, #533, #536).
Memory use is lower: the pass-1 feature table shrank to 42 bytes per row, and the density table and line-joining step use less memory (#543, #565, #570, #579, #581, #626).
Zoom levels that share a simplification step compute it once (#499, #525).

Internal¶

Docs rebuilt around two tested tutorials (#639). Release builds check for integer overflow (#432, #589). Line clipping now uses i_overlay 9 only (#205, #435, #597). New regression guards, benchmarks against tippecanoe, sharded slow tests, dependency updates and CI work (#447, #448, #524, #558, #562, #566, #567, #612, #619).

v0.7.1 (2026-09-23)¶

Packaging and documentation updates for v0.7.0, with easier installation and a new quickstart. The tiling engine is unchanged.

Added¶

Python wheels: pip install tylertoo uses prebuilt wheels on Linux x86_64 and aarch64 (manylinux and musllinux, including Alpine), macOS Intel and Apple Silicon, and Windows x86_64. Each stable-ABI wheel (abi3-py39) supports CPython 3.9 and newer without source builds (#410, #473).
cargo binstall tylertoo downloads a prebuilt release binary without requiring Rust or protoc (#411, #469).
60-second quickstart in the README and at the start of the getting-started guide, using a downloadable 28 MB sample file. Every command was tested (#413, #472).
Community files: SECURITY.md with private vulnerability reporting, a code of conduct, issue forms, a PR template, and issues labelled good-first-issue (#415, #467).

Changed¶

Reproducible builds: Cargo.lock is committed, and release binaries, wheels and crates.io publishes use --locked to build from reviewed dependencies. v0.7.0 was built without a lockfile (#412, #468).
Clearer README intro and a curated v0.7.0 changelog. The Brazil tutorial now links to the 52-file remote manifest; the previously named single file never existed. The gpio commands no longer use a flag absent from gpio 1.5.0 (#414, #472).

Fixed¶

Release pipeline: natively built wheels are smoke-tested (installed and imported) before publication. Manual workflow re-runs cannot upload wheels under an already released version (#473).

Internal¶

uv.lock version-sync hook and CI check (#466); geometry-test-data submodule setup documented (#464, #474); gpq-tiles tombstone runbook (#470); rust-toolchain action re-pin (#475); dependency and action bumps (#477, #478).

v0.7.0 (2026-09-19)¶

Replaces the original per-tile pipeline with geo:overviews: build a multi-resolution overview GeoParquet, then export tiles from it. The overview file can be validated, queried with SQL, and re-exported.

Added¶

geo:overviews GeoParquet overviews — tylertoo overview embeds COG-style multi-resolution levels in a single valid GeoParquet file; tylertoo validate checks one against the draft spec (#168, #184, #190).
tylertoo export-pmtiles — PMTiles v3 export from an overview file, with per-level progress logging and incremental checkpoints (#169, #229).
tylertoo tiles, and the bare form — one-shot GeoParquet → PMTiles via overview → export, with --keep-overview, a configurable spill directory, a free-space preflight, and the same tuning options as the two-step commands (#251, #276, #318, #319).
tylertoo decode — PMTiles v3 → GeoParquet with tippecanoe-decode semantics and zoom / layer / mvt_id provenance columns (#112, #206).
tylertoo pyramid — merge several inputs with disjoint zoom ranges into one PMTiles archive. Bands may be GeoParquet (tiled during conversion) or existing PMTiles. Bands in different layers may share a zoom range (#348, #385, #392).
Remote input — read s3://, https:// and gs:// GeoParquet over byte-range requests, including remote prefix listing (#210, #216, #281).
Multi-file input — a local directory, a glob, an s3:// / gs:// prefix, a --files-from manifest (ordered, mixed local and remote), or a Python list[str] is read as one dataset, with cross-partition schema and CRS validation and a deterministic row order (#277, #281, #282).
--spill-dir and a free-space guard — choose where to store the remote-input spill file. Before pass 1, the converter estimates its size and warns about potentially insufficient space or tmpfs storage (#272, #273). Conversions that would stage roughly the whole remote object also show equivalent download-then-convert commands up front (#267).
--bbox regional extracts — spatial pushdown skips row groups outside the requested region, avoiding unnecessary reads and remote fetches (#102, #207).
--filter / --where — SQL-WHERE-style attribute predicates with parquet row-group statistics pushdown, timestamp columns included (#315, #321).
Property selection — --include-property, --exclude-property and --exclude-all-properties (tippecanoe -y / -x / -X), applied at scan time on overview and at encode time on export-pmtiles (#386, #391).
--feature-order — set within-tile draw order to input order or a property, ascending or descending (#361, #366).
Magnitude ladder — --magnitude-ladder and --entry-zoom set each feature's first zoom from an attribute (#364, #375).
--verbatim — tile the input as given, disabling generalization at every level (#345, #360, #367).
Zoom-band representation — --representation "0-7:point" draws a band as representative points and the rest as polygons in one archive, and --collapse-square replaces dropped tiny polygons with area-dithered placeholder squares (#279, #317, #322).
Tiny-polygon accumulator — preserve the area of dropped polygons at coarse levels (#384, #394).
Per-level point clustering — --cluster and --accumulate-attribute carry cluster winners with a point_count column and numeric attribute aggregation through the overview pipeline (#170).
Line coalescing — on by default, with --coalesce-snap, --coalesce-max-level-rows and --no-coalesce-lines (#183).
Cascading simplification — each level simplifies the one above it rather than the source; --no-cascade opts out (#218).
--report — machine-readable JSON run reports from overview, export-pmtiles and tiles (#316, #318).
Python overview API — overview(), validate() and export_pmtiles() bindings, with convert() kept as a facade over them (#185).
Prebuilt CLI binaries attached to every GitHub Release — Linux x86_64 (gnu and musl), macOS Intel and Apple Silicon, Windows x86_64 — plus generated release notes (#275).
Warn during conversion about features suspected of crossing the antimeridian (#188, #199).

Changed¶

Breaking: the legacy per-tile pipeline is gone. Removed pipeline.rs, Converter, TilerConfig and their library re-exports. All conversion now runs overview → export (#177, #189). Removed the old pipeline flags: --streaming-mode, --include / --exclude / --exclude-all property filtering, output-compression selection, and --deterministic. Use the overview / export-pmtiles tuning flags, also accepted by tiles (#249).
Breaking: Python convert() is deprecated. It no longer runs the removed legacy pipeline; it chains overview() + export_pmtiles(). Use the two-step API for all options. Its legacy keyword arguments drop_density, compression, include, exclude, exclude_all, deterministic, drop_smallest_as_needed, drop_smallest_threshold and progress_callback were removed; passing one raises TypeError.
--polygon-visibility retuned 4.0 → 2.0 (#259): the rendered sweep showed gates above 2.0 remove too many features at coarse zooms without reducing file size. Gates below ~2.0 mostly admit candidates dropped by write-time collapse (corpus/SWEEPS.md, Decision 6).
Per-tile size caps at 500K by default, matching tippecanoe (#280).
The simple-clip fast path is on by default; --no-simple-clip-fastpath opts out (#239, #256).
Line thinning retuned 2.0 → 1.0, point thinning defaults to 16.0 under clustering, and junction continuation now defaults off, based on the same sweep methodology.
Overview footers declare spec version 0.2.0 (#184, #190).
The declared MSRV is now 1.95, corrected from 1.75 and verified in CI (#358).

Fixed¶

PMTiles directories: the contiguous-offset rule now applies to leaf entries, oversized directories split into leaves, and leaf_dirs_offset is never written as 0 (#356, #377, #378).
MVT polygons are cleaned in tile space after quantization, preventing self-intersecting or degenerate rings (#383, #393). Ring winding follows the spec.
export-pmtiles declares the requested minimum zoom even when the coarse levels generalized to nothing (#380, #390).
coalesced_count is withheld from tiles when it is 1 everywhere (#379, #389).
Renamed source columns are published under their source name (#359, #365).
DATE, TIMESTAMP and DECIMAL properties are encoded instead of dropped, with timestamps UTC-stamped.
--tile-buffer is read as tile pixels, not MVT extent units, and tile membership derives from the buffer-expanded bbox.
MultiPolygon parts that split into pieces at a tile boundary are kept (#244).
Empty overview levels auto-clamp instead of failing the run (#211).
Reserved-column collisions are auto-renamed rather than rejected (#288).
An explicit null GeoParquet crs is assumed to be OGC:CRS84.
Coarse-level export used geographic latitude instead of Web Mercator Y in the tile-local transform, distorting high-latitude tiles.
An input with a case-insensitive level column collision is rejected up front instead of producing an ambiguous overview file.
decode names the section or tile that failed to decompress, with its byte offset (#381, #388).
Oversized remote column chunks are no longer re-fetched per page (#261), and http:// inputs work (#262).

Performance¶

Remote-input network traffic is bounded to about 1× the object's bytes, down from ~3×, regardless of zoom range or pass count. Selected row groups are staged up front for coalesced, parallel fetches; column chunks are cached to the largest row group's working set (#261); and an on-disk spill file serves passes 2 and later (#219, #286, #287). If the spill volume fills, conversion continues by re-fetching from the network.
Export uses a single-read fan-out pass to avoid cross-level prefix and per-partition re-reads, with parallel clipping, simplification, MVT encoding and gzip (#227, #228, #233, #235).
Large polygons split top-down recursively instead of being clipped whole (#226), and three O(V²) hot spots in clipping and validity checking were replaced or capped (#237, #241, #242).
Memory budgeting uses winner-grid level waves, partition waves sized from the densest partition and available cores, and an auto profile that selects RAM or spill storage from measured geometry (#293, #294, #303, #305, #306, #311).
Parquet row-group encoding and GeoParquet WKB encoding run in parallel (#296, #304).
--profile presets over a single-read pipelined pass-2 engine (#212, #213).

Internal¶

For Clippy, rustfmt, dependency, CI and benchmark-harness changes, see the full 465-commit record: git log v0.6.0..v0.7.0.

v0.6.0 (2026-03-11)¶

Feat¶

implement point clustering with position averaging (#25)
implement accumulator system for attribute aggregation (#23)
implement gap-based density detection (#24)
support WKT geometry encoding (#35)
implement tiny polygon accumulation (#85)
profiling: add fine-grained spans to read_parquet phase
add time profiling with tracing
add memory profiling with dhat

Fix¶

remove needless borrow in benchmark
resolve clippy warnings and unused imports
remove WKT fixture from repo, tests skip when missing
gracefully skip WKT tests when fixture is missing
skip profiling integration tests when dhat-heap feature enabled
use tempfile crate for proper temp directory isolation in tests
use cross-platform temp directories in integration tests

Refactor¶

replace wagyu-rs with i_overlay for polygon clipping

Perf¶

parallel row group I/O for ~24% speedup
reuse file handle across row groups (#41)
parallelize tile encoding in Phase 3 (#90)

v0.5.0 (2026-03-10)¶

Feat¶

core: wire up pipeline to use WorldCoord throughout (Phase 2)
core: add WorldCoord-based hierarchical clipping (Phase 2)
core: add WorldCoord-based feature drop functions (Phase 2)
core: add WorldCoord support to MVT encoding and validation (Phase 1)
core: add WorldCoord support to clipping modules (Phase 1)
core: add WorldCoord-based simplification functions (Phase 1)
core: add WorldCoord type for 32-bit integer coordinates (Phase 0)
clip: integrate wagyu-rs for robust polygon clipping
add --deterministic flag and fix PR #63 review feedback

Fix¶

clip: enable U-shape split test with wagyu-rs v0.2.1
clip: add wagyu fallback for edge case geometry handling (#94)
change default compression to gzip and add CRS validation
change default compression to gzip for compatibility
implement leaf directory support for large PMTiles archives
core: clamp tile coordinates and bounds to valid ranges
core: fix issue #83 - geometry coordinates collapsing to zeros
resolve clippy warnings in tests
core: align feature_drop coordinate precision with MVT encoding
ci: download fixtures from release instead of LFS
ci: remove 1.8GB fixture from LFS to fix bandwidth quota
Remove unused MIN_EXPECTED_TILES constant
Remove #[ignore] from regression tests - fixture is in LFS
use is_empty() instead of len() > 0 for Clippy
Fix clippy warning and add clipping benchmarks
tile: clamp latitude to Web Mercator bounds
use wagyu-rs from crates.io instead of path dependency
ci: update benchmark group names after consolidation

Perf¶

Add pre-clip bounding box filter for large geometries
Implement hierarchical clipping across zoom levels
Replace Wagyu with Sutherland-Hodgman for tile clipping

v0.4.0 (2026-02-25)¶

Feat¶

python: add progress callback support
python: add streaming mode and parallel control parameters
python: add property filtering and layer name parameters

Fix¶

add version sync safeguards and fix pyproject.toml version mismatch
add clippy allow for too_many_arguments and add clippy to pre-commit
use workspace dependency for gpq-tiles-core version

v0.2.0 (2026-02-25)¶

Fix¶

release: complete v0.2.0 release setup

v0.1.0 (2026-02-24)¶

Feat¶

set up commitizen for automated versioning and releases
quality: warn about pathologically small row groups
default to zstd compression, expose parallel options in CLI
add progress bars for cleaner output
parallelize geometry processing within row groups (#37)
parallelize tile processing for large geometries (#33)
cli: add --streaming-mode flag with progress reporting
pipeline: implement ExternalSort streaming mode
pipeline: add StreamingMode::ExternalSort variant
core: add external sort and WKB serialization modules
streaming: add StreamingPmtilesWriter with LowMemory mode
streaming: add memory budget configuration and tracking
streaming: add row-group-based streaming tile generation
quality: add GeoParquet file quality detection for streaming
add tile deduplication with XXH3 hashing and run_length encoding
add compression options (gzip, brotli, zstd, none) for PMTiles output
add property filtering with --include/-y, --exclude/-x, --exclude-all/-X flags
add 17K feature fixture for parallelization benchmarks
add tilestats metadata to PMTiles output
auto-extract field metadata from GeoParquet schema
add field metadata support to PMTiles writer
derive layer name from input filename, add --layer-name CLI flag
complete Phase 5 Python bindings with uv/ruff tooling
add benchmark suite with generate_tiles_from_geometries API
pipeline: add Rayon parallel tile generation
pipeline: wire spatial indexing into tile generation
spatial-index: add space-filling curve sorting for efficient tile generation
complete Phase 3 with density-based dropping
integrate feature dropping into pipeline (Phase 3)
add point thinning (1/2.5 drop rate per zoom)
add line dropping (coordinate quantization algorithm)
implement tiny polygon dropping with diffuse probability
implement PMTiles v3 writer (Tasks 7-9)
implement tiler pipeline wiring clip → simplify → MVT
implement MVT encoding for vector tiles
add golden comparison tests against tippecanoe output
implement geometry clipping with correct BooleanOps
implement zoom-based simplification
implement Arrow-native geometry batch processing (TDD green)

Fix¶

release: add version to core dep, add READMEs, fix benchmark filter
release: install protoc inside manylinux container
python: copy README into crate for sdist builds
ci: use tag triggers for release, skip slow benchmarks
consolidate release workflows and fix version bump detection
guard against degenerate linestrings in simplify and fix flaky test
PMTiles now compatible with pmtiles.io and standard viewers
golden: update stale Z8 test to use full pipeline
resolve three medium/low priority issues
simplify geometry in tile-local pixel coordinates
handle antimeridian crossing in tiles_for_bbox
clip: preserve all polygon parts when clipping produces MultiPolygon
use real bbox calculation instead of world bounds
resolve CI timeouts and coverage linker errors
upgrade pyo3 0.24 → 0.28 for Python 3.14 support
resolve CI failures for benchmark, check, test, and security audit

Perf¶

streaming: add memory benchmarks for streaming pipeline