A performant binary encoding for geographic data based on flatbuffers that can hold a collection of Simple Features.
Inspired by geobuf and flatbush. Deliberately does not support random writes for simplicity and to be able to cluster the data on a packed Hilbert R-Tree enabling fast bounding box spatial filtering. The spatial index is however optional to allow the format to be efficiently written as a stream.
Goals are to be suitable for large volumes of static data, significantly faster than legacy formats without size limitations for contents or metainformation and to be suitable for streaming/random access.
The site http://switchfromshapefile.org has more in depth information about the problems of legacy formats and provides some alternatives but acknowledges that the current alternatives has some drawbacks on their own, for example they are not suitable for streaming.
Live demonstration at https://observablehq.com/@bjornharrtell/streaming-flatgeobuf. (conceptual, not performance optimized)
- MB: Magic bytes (0x6667620066676200)
- H: Header (variable size flatbuffer)
- I (optional): Static packed Hilbert R-tree index (static size custom buffer)
- O: Feature offsets index (static size custom buffer, feature count * 8 bytes)
- DATA: Features (variable size flatbuffers)
Any 64-bit flatbuffer value contained anywhere in the file (for example coordinate values) is aligned to 8 bytes to from the start of the file to allow for direct memory access.
Preliminary performance tests has been done using road data from OSM for Denmark in SHP format from https://download.geofabrik.de/, containing 812547 LineString features with a set of attributes.
| Shapefile | GeoPackage | FlatGeobuf | GeoJSON | GML | |
|---|---|---|---|---|---|
| Read full dataset | 1 | 0.9 | 0.5 | 15 | 7.7 |
| Read w/spatial filter | 1 | 0.15 | 0.12 | 100 | 60 |
| Write full dataset | 1 | 0.62 | 0.37 | 2.5 | 2 |
| Write w/spatial index | 1 | 1.3 | 0.45 | - | - |
The test was done using GDAL implementing FlatGeobuf as a driver and measurements for repeated reads using loops of ogrinfo -qq -oo VERIFY_BUFFERS=NO runs and measurements for repeated writes was done with ogr2ogr -oo VERIFY_BUFFERS=NO conversion from the original to a new file with -lco SPATIAL_INDEX=NO and -lco SPATIAL_INDEX=YES respectively.
Note that for the test with spatial filter a small bounding box was chosen resulting in only 9 features. The reason for this is to test mainly the spatial index search performance for that case.
- Language support for JavaScript, TypeScript, C, C++, Java and C#
- Efficient I/O (streaming and random access)
- GDAL/OGR driver
- QGIS provider (WIP @ https://github.com/bjornharrtell/QGIS/tree/fgb)
- OpenLayers example (WIP @ https://github.com/bjornharrtell/ol3/tree/flatgeobuf)
- GeoServer WFS output format (WIP @ https://github.com/bjornharrtell/geoserver/tree/flatgeobuf-output)
- Complete test coverage
- Java index support
- C# support update
- C langauge support
- Go langauge support
- Rust language support
- Further optimizations
It does not align on 8 bytes so it not always possible to consume it without copying first.
Performance reasons and to allow streaming/random access.
Allowing per feature schema breaks the simple in simple features, in my opinion.
Same reason as to why I prefer the static schema requirement.
Default behaviour is to assume untrusted data and verify buffer integrity for safety. If you have trusted data and want maximum performance make sure to set the open option VERIFY_BUFFERS to NO.