Skip to main content

manby

manby builds vector tilesets from GeoJSON, FlatGeobuf, CSV, Geobuf, GeoParquet or Arrow features, GeoParquet in object storage, and from DuckDB queries and data frames in Python and R. It makes the same tiles as tippecanoe, in less memory, ten to eighteen times faster on a Mac and eight to fifteen times faster on Linux, more where tippecanoe's disk slows it down. manby join is tippecanoe's tile-join, two to thirty-five times faster.

The manby command takes tippecanoe's options:

manby -zg --drop-densest-as-needed -o places.pmtiles places.geojson

pip install manby installs it on macOS and Linux, with a Python module, and uvx manby runs it without installing. brew install birkskyum/tap/manby and mise use -g github:birkskyum/manby install the program alone from the releases. cargo install manby builds it, with Rust 1.89 or later. The site has the guides.

From Python, a DuckDB query streams to manby in Arrow batches.

import duckdb, manby

places = duckdb.sql("SELECT name, geometry FROM 'places.parquet' WHERE country = 'DE'")
manby.tile(places, "places.pmtiles", maximum_zoom="g", drop_densest_as_needed=True)

The R package, from R-universe, tiles sf data frames and DuckDB tables the same way. manby serve serves tilesets with a map of them, and manby stats reports their tiles' sizes, from the command line, R or Python.

manby::tile(sf::st_read("places.gpkg"), "places.pmtiles", maximum_zoom = "g", drop_densest_as_needed = TRUE)

Speed

Against tippecanoe 2.82.0 itself, the C++ built with its own Makefile, writing PMTiles. Each figure is the median of three runs, and memory is peak resident memory. tippecanoe's runs and manby's come from separate sessions on the same machine, since tippecanoe's do not change with manby's version.

On macOS, on an Apple M5 Max with 18 cores and 128 GB:

Benchmark tippecanoe manby
4 million points, -zg --drop-densest-as-needed 61.4 s, 2,276 MB 3.44 s, 919 MB 17.8x
The same points, -zg -r1 --cluster-distance=10 --cluster-densest-as-needed --accumulate-attribute=value:sum 54.4 s, 1,318 MB 3.88 s, 891 MB 14.0x
The same points, -zg --cluster-densest-as-needed 59.8 s, 2,330 MB 4.07 s, 916 MB 14.7x
The same points, -zg -r1 --cluster-densest-as-needed --accumulate-attribute=value:sum 87.7 s, 3,443 MB 7.82 s, 1,115 MB 11.2x
17,522 roads at -z14, 122,725 tiles 11.8 s, 327 MB 0.81 s, 177 MB 14.5x
85,186 census tracts at -z12, 312,073 tiles 46.6 s, 566 MB 3.78 s, 338 MB 12.3x
85,186 census tracts, -zg --coalesce-densest-as-needed 24.8 s, 554 MB 2.27 s, 366 MB 10.9x

On Linux, on a Lenovo Yoga Slim 7x with a Snapdragon X Elite (12 cores, 32 GB), running Arch Linux ARM:

Benchmark tippecanoe manby
4 million points, -zg --drop-densest-as-needed 88.5 s, 1,088 MB 5.86 s, 994 MB 15.1x
The same points, clustered within 10 pixels, summing value 82.5 s, 1,028 MB 6.02 s, 974 MB 13.7x
The same points, -zg --cluster-densest-as-needed 89.6 s, 1,165 MB 6.37 s, 1,000 MB 14.1x
The same points, -zg -r1 --cluster-densest-as-needed, summing value 133.6 s, 1,498 MB 13.9 s, 1,152 MB 9.6x
17,522 roads at -z14, 122,725 tiles 59.0 s, 240 MB 1.51 s, 236 MB 39.1x
85,186 census tracts at -z12, 312,073 tiles 196.8 s, 456 MB 7.77 s, 360 MB 25.3x
85,186 census tracts, -zg --coalesce-densest-as-needed 36.1 s, 451 MB 4.46 s, 379 MB 8.1x

On this machine tippecanoe uses 8 of the 12 cores, since it rounds its thread count down to a power of 2, which keeps its memory low, while manby uses all 12. tippecanoe is slow on the runs with many tiles here because it writes PMTiles by way of a SQLite database next to the output, on a btrfs filesystem that compresses with zstd. With the output in memory instead, roads take it 27 s rather than 59. The kernel also backs memory with 2 MB pages wherever it can (transparent huge pages set to always, as Arch sets them), which adds about a tenth to manby's figures, and leaves its roads only just below tippecanoe's. bench/ has the data, the scripts and how the Linux numbers were made.

For every run the tiles decode to exactly what tippecanoe makes, with one exception. tippecanoe's arm64 builds, from Apple's clang on macOS and from GCC on Linux, fuse multiplications and additions where its x86-64 build does not, and that changes a few tiles near the size limit. On these benchmarks it changes only the census tracts at -zg, here and 20 times over below. In the tracts, the zoom 0 tile comes out 6 bytes smaller compressed, so that tippecanoe's retry keeps other features. manby makes the same tiles on macOS and Linux alike, the tiles of tippecanoe's x86-64 build. Built for x86-64, or on arm64 with -ffp-contract=off, tippecanoe makes manby's tiles on the tracts too.

tippecanoe-rs, the faithful Rust port that manby takes tippecanoe's algorithms from, runs at about the speed of the C++. So porting tippecanoe to Rust is not what makes manby fast. Doing the work differently is.

The archives are not tippecanoe's byte for byte, and a little smaller. zlib's exact output decides which features a tile keeps only when the tile could be near the size limit. Every other tile is compressed with libdeflate, which is faster and compresses tiles better. --compression=zlib compresses every tile with zlib, as tippecanoe does, for the same archive bytes.

How

tippecanoe spends most of its time waiting on itself rather than computing. It works through the low zoom levels one tile per thread, rebuilds oversized tiles by reading their features back from compressed temporary files, finishes each zoom level before starting the next, and writes PMTiles by way of a SQLite database. manby keeps the features in memory, sorted once, walks the tile pyramid in parallel on every core, splits big tiles across threads, retries oversized tiles on data already in memory, and writes PMTiles directly. The geometry algorithms are tippecanoe's own, by way of tippecanoe-rs, which is why the tiles come out the same. DESIGN.md has the measurements and the architecture.

Memory

manby's peak memory is below tippecanoe's on the benchmarks. For inputs larger than memory, --memory-budget bounds it (default: half the physical memory, as tippecanoe sizes its sorting). Past the budget, the features and the tile data live in an unlinked temporary file in the -t directory, mapped into memory, laid out in the order the tiling reads them, and manby keeps its resident memory near the budget. bench/large.sh tiles inputs several times larger than a 1 GB budget, once each:

Input System tippecanoe manby, --memory-budget=1G
40 million points (5.6 GB), -zg --drop-densest-as-needed macOS 652 s, 10,206 MB 79 s, 1,139 MB
Linux 1,025 s, 6,379 MB 121 s, 1,188 MB
The census tracts 20 times over (3 GB), -zg --coalesce-densest-as-needed macOS 564 s, 4,810 MB 137 s, 1,022 MB
Linux 854 s, 3,730 MB 269 s, 1,206 MB

Joining tilesets

manby join takes tile-join's options, prints its messages, exits as it does, and makes the same tiles, with their features in the same order. It merges tilesets, joins attributes from a CSV file, filters and renames layers, filters features, overzooms, and converts between MBTiles, PMTiles and directories:

manby join -o merged.pmtiles roads.pmtiles places.mbtiles
manby join -c tracts.csv -o joined.pmtiles tracts.pmtiles

manby tile-join works too. Under the name tile-join, a link to it for example, manby is tile-join, so it can take its place.

On the same machines, median of three runs, against tile-join 2.82.0, the C++. For every run the tiles decode to exactly what tile-join makes.

Benchmark System tile-join manby join
The tracts at -z12 (312,073 tiles) copied to PMTiles macOS 21.9 s, 589 MB 0.63 s, 161 MB 34.8x
Linux 24.2 s, 461 MB 1.05 s, 185 MB 23.1x
The same, to MBTiles macOS 16.6 s, 560 MB 1.19 s, 380 MB 13.9x
Linux 20.4 s, 463 MB 1.28 s, 371 MB 16.0x
The same tracts joined to a CSV keyed on GEOID macOS 22.6 s, 652 MB 1.09 s, 296 MB 20.7x
Linux 25.9 s, 556 MB 1.92 s, 325 MB 13.5x
The points and the tracts, both -zg, merged to PMTiles macOS 1.20 s, 701 MB 0.38 s, 269 MB 3.2x
Linux 1.62 s, 573 MB 0.68 s, 249 MB 2.4x
The same, to MBTiles macOS 1.15 s, 707 MB 0.40 s, 273 MB 2.9x
Linux 1.51 s, 567 MB 0.69 s, 258 MB 2.2x
Counties at -z5 and tracts at -z9, with --overzoom macOS 3.87 s, 592 MB 0.39 s, 182 MB 9.9x
Linux 5.99 s, 440 MB 0.72 s, 196 MB 8.3x

The merges spend most of their time on a hundred tiles near the 500 KB limit, which manby join, like tile-join, has to compress with zlib to check their size. bench/join.sh has the benchmarks.

tile-join decodes each tile into layers that each allocate a megabyte of hash tables, does its overzooming one tile at a time, compresses every tile with zlib, and writes PMTiles by way of a SQLite database. manby join keeps tile-join's order of input tiles and its metadata, does the work on the tiles on every core without those tables, keeps the compression of tiles it does not change, and writes PMTiles directly. DESIGN.md has the details.

Where several inputs have a tile at the same z/x/y, tile-join appends their features in the order of the tiles' stored bytes. manby compresses most tiles with libdeflate, whose bytes sort differently from zlib's, so for tilesets that manby made, manby join compares tiles by the bytes zlib makes of them, which are tippecanoe's. So manby and then manby join make the tiles that tippecanoe and then tile-join make, with the tiles of tippecanoe's x86-64 build as above. tile-join itself, run on manby's tilesets, can order the features of those shared tiles differently. Tilesets that manby made with --compression=zlib are joined the same way by both.

tippecanoe's other programs

manby also runs tippecanoe's smaller programs, as tippecanoe-rs has them, so one program does all of tippecanoe's work:

tippecanoe manby
tippecanoe manby
tile-join manby join
tippecanoe-decode manby decode
tippecanoe-enumerate manby enumerate
tippecanoe-overzoom manby overzoom
tippecanoe-json-tool manby json-tool

Each program's own name works as the subcommand too (manby tippecanoe-decode), and under that name, through a link, manby is that program, so links named after all six make manby a drop-in replacement for tippecanoe. Only the tiling and joining are manby's own, since that is where the time goes. The other four are tippecanoe-rs's code and run at its speed.

To tile an input file named like a subcommand, such as decode, write it as ./decode.

Building and testing

manby uses tippecanoe-rs, in crates/tippecanoe-rs, as a library. The port has tippecanoe's six programs, built with cargo build --release -p tippecanoe-rs, and runs tippecanoe's test suite against them with make -C crates/tippecanoe-rs test.

cargo build --release
make test

make test runs the unit tests and tippecanoe's own test suite with manby in place of tippecanoe, from GeoJSON input and again from Geobuf input. All 188 of tippecanoe's expected outputs match, and the tests of tippecanoe's other programs pass with manby as each of them.

manby is a library first. manby::run(&Options) reads, tiles and writes a tileset and returns an error instead of exiting, manby::join(&JoinOptions) joins tilesets, and the commands are thin layers over them.

MapLibre Tiles

--tile-format=mlt writes MapLibre Tiles instead of Mapbox Vector Tiles, with the same features, into PMTiles, MBTiles or a directory. The tiles are 2 to 45% smaller than MVT on the benchmarks, and MapLibre GL JS 5.12 and later reads them from a vector source with encoding: "mlt".

manby -zg --drop-densest-as-needed --tile-format=mlt -o places.pmtiles places.geojson

The wheels and the release binaries write MLT. The encoder, mlt-core, needs Rust 1.98, so from source MLT output is a cargo feature, cargo +1.98.1 build --release --features mlt.

DESIGN.md describes the choices behind the tiles, and make test-golden-mlt checks them against the MVT tiles of the same runs.

Fitting each tile on its own

When a tile is too big, tippecanoe raises its threshold for dropping, coalescing or clustering until it fits, then makes the whole zoom level again with the highest threshold any tile needed. --as-needed-per-tile keeps each tile's own threshold instead. Only the tiles that are too big are thinned, so the others keep more, but neighboring tiles can be thinned differently, which can show at tile edges. On the census tracts, the tiles of z5 keep 67,163 tract polygons rather than 17,094, and the run is 1.2 times faster. The tiles are not tippecanoe's.

Thanks

manby is built on tippecanoe. Erica Fischer started it in 2014 and, over twelve years at Mapbox and then at Felt, wrote nearly all of it. The way it thins features evenly across a zoom level until every tile fits, and the test suite that pins all of that down, are what manby keeps, and they are her work. Thank you, Erica.

License

BSD 2-Clause, see LICENSE.md.

Metadata

Release files for manby 0.1.7

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Built distributions (wheels)

Table of built distributions (wheels) for manby 0.1.7
File Actions
manby-0.1.7-py3-none-musllinux_1_2_x86_64.whl Python 3 none Linux musl 1.2+ x86-64 Details for manby-0.1.7-py3-none-musllinux_1_2_x86_64.whl
manby-0.1.7-py3-none-musllinux_1_2_aarch64.whl Python 3 none Linux musl 1.2+ ARM64 Details for manby-0.1.7-py3-none-musllinux_1_2_aarch64.whl
manby-0.1.7-py3-none-manylinux_2_28_x86_64.whl Python 3 none Linux glibc 2.28+ x86-64 Details for manby-0.1.7-py3-none-manylinux_2_28_x86_64.whl
manby-0.1.7-py3-none-manylinux_2_28_aarch64.whl Python 3 none Linux glibc 2.28+ ARM64 Details for manby-0.1.7-py3-none-manylinux_2_28_aarch64.whl
manby-0.1.7-py3-none-macosx_11_0_arm64.whl Python 3 none macOS 11.0+ ARM64 Details for manby-0.1.7-py3-none-macosx_11_0_arm64.whl
manby-0.1.7-py3-none-macosx_10_12_x86_64.whl Python 3 none macOS 10.12+ x86-64 Details for manby-0.1.7-py3-none-macosx_10_12_x86_64.whl

Total release size: 57.4 MB

Release files / manby-0.1.7-py3-none-musllinux_1_2_x86_64.whl

Download URL manby-0.1.7-py3-none-musllinux_1_2_x86_64.whl
Size 10.2 MB
Tags Linux musl 1.2+ x86-64 Python 3
SHA-256 checksum
How to use checksums
616794f06f38a69768227bb8bf15e65f7099c97398a55a13825667466c9a9c57
BLAKE2b-256 checksum
How to use checksums
efea919e2aadb01d86fb138a54e3a2e0bb4e0848e7087190f3cf5c875b5ecace
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 8, 2026.

Transparency log

Release files / manby-0.1.7-py3-none-musllinux_1_2_aarch64.whl

Download URL manby-0.1.7-py3-none-musllinux_1_2_aarch64.whl
Size 9.5 MB
Tags Linux musl 1.2+ ARM64 Python 3
SHA-256 checksum
How to use checksums
90742e28fe5bdaab9da87e1709975ca78ababe9d2c8d65194888ac4b157f8fb3
BLAKE2b-256 checksum
How to use checksums
6cd9312ead0606f153dc3fd640c81aceeaf6c177c021d96da148c99672e35212
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 8, 2026.

Transparency log

Release files / manby-0.1.7-py3-none-manylinux_2_28_x86_64.whl

Download URL manby-0.1.7-py3-none-manylinux_2_28_x86_64.whl
Size 10.1 MB
Tags Linux glibc 2.28+ x86-64 Python 3
SHA-256 checksum
How to use checksums
dbef07a38da80ec4064ba9ee08c76929f678585fabc1a2703b3b9966c7209982
BLAKE2b-256 checksum
How to use checksums
7a5573facfc1f0c103b64a844512f9aa8bcbb2ae0bac9003594ad8d3ba4e6556
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 8, 2026.

Transparency log

Release files / manby-0.1.7-py3-none-manylinux_2_28_aarch64.whl

Download URL manby-0.1.7-py3-none-manylinux_2_28_aarch64.whl
Size 9.3 MB
Tags Linux glibc 2.28+ ARM64 Python 3
SHA-256 checksum
How to use checksums
4cde228d4fc189ad0e94e693a4248b66151453ba145aae227fbe76173b77b61b
BLAKE2b-256 checksum
How to use checksums
35661135cd71d665f4922b07b19e59cfe1bf4bdfa835cbcd5627d579c3a8344d
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 8, 2026.

Transparency log

Release files / manby-0.1.7-py3-none-macosx_11_0_arm64.whl

Download URL manby-0.1.7-py3-none-macosx_11_0_arm64.whl
Size 8.5 MB
Tags Python 3 macOS 11.0+ ARM64
SHA-256 checksum
How to use checksums
863efa714ab428c9b97d6a987a873434446b3664d9b67dc6ae0f9428d63e6c78
BLAKE2b-256 checksum
How to use checksums
d22251629e2c1340541d108d5e1587cab3472ccfc0440bb6dea157008c199262
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 8, 2026.

Transparency log

Release files / manby-0.1.7-py3-none-macosx_10_12_x86_64.whl

Download URL manby-0.1.7-py3-none-macosx_10_12_x86_64.whl
Size 9.7 MB
Tags Python 3 macOS 10.12+ x86-64
SHA-256 checksum
How to use checksums
da471f79840c6cba4b6a88b80c2e0d3671809f2900b785da4df0484ecc707068
BLAKE2b-256 checksum
How to use checksums
152e53a6f616daca5a40576ce785e013e86b2d5409a280fed7e0630155f26a1e
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 8, 2026.

Transparency log

Security for manby 0.1.7

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

All 6 files in this release have been uploaded with provenance metadata.

Select from 6 files

Attestation for manby-0.1.7-py3-none-musllinux_1_2_x86_64.whl

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 8, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page