GLX specification changelog | geolith .glx docs

Generated from spec/v1/CHANGELOG.md in the geolith/glx repository at commit 561a984.

1.0.3

A routing part may hold a customizable contraction hierarchy per costing profile (10.5). Routing parts without one are unchanged.

Feature bit 0 (CCH): the routing header grows to 80 bytes with the CCH directory's (offset, length); the directory ends the boot span.
Per profile, three blocked sections after the tiles: the vertex map (edge to ranks and ancestor runs), the records (kept up and down directions with weights in 0.1 cost units) and the road lists.

The routing part's layout changes; this is still version 1 (magic GLX0, version 1). Routing parts written to 1.0.2 do not read: rebuild them. Search parts are unchanged.

Graph tiles are stored as fixed-size blocks, each its own zstd frame, behind a per-tile block directory, so a reader fetches and holds only the blocks a route touches (10.2, 10.3).
The routing header's former reserved field at offset 4 is the block size, a power of two from 4 KiB to 16 MiB; the reference writer uses 64 KiB (10.4).
Tile index entries grow to 16 bytes: key, tile length, first frame length, stored length.
Chapter 10.4 records the Berlin measurements behind the block size.

The file layout changes; this is still version 1 (magic GLX0, version 1). Files written to 1.0 or 1.0.1 do not read: rebuild them. A reader for 1.0.2 refuses them, because their old header does not hold a valid part table.

Every file starts with a 256-byte file header with a part table of up to seven entries: kind, offset, length and XXH3-64 checksum (2.6).
The search data is the search part, kind SRCH. Its 256-byte header and sections are the former file layout; every offset in it now counts from the start of the part (Chapter 3).
A routing part, kind ROUT, holds road graph tiles as separate zstd frames with a tile index and an optional dictionary (Chapter 10).
A reader for one kind of part refuses a file that holds no part of that kind, and ignores parts of other kinds.
A tool that copies a part between files checks its checksum; the parts of several files merge into one file by copying each part's bytes unchanged.

1.0.1

The file layout is unchanged: header version 1, magic GLX0, and every file written to 1.0 reads the same.

Readers MUST refuse a file that sets a reserved feature bit (17 to 31), rather than MAY ignore it. A bit a later revision adds is thus refused by older readers, required or optional.
Readers MUST refuse a term entry of a reserved kind (above 2), rather than SHOULD.
Readers MUST refuse a header whose document count exceeds 2^32 - 1.
Latitudes clamp to ±90 degrees in the fixed-point encoding; longitudes still clamp to ±180. Real latitudes lie within ±90, so written files do not change.
The magic and the version field are explained: the magic names the format and stays GLX0; the version field names the specification.
Feature bits 8 and 9 are stated to be set even when their sections are empty.

1.0

Initial version of the GLX format: header version 1, magic GLX0.

A 256-byte header with thirteen sections: term index, spatial directory, document index, chunks, dictionary, postings, documents, spatial, id directory, ids, shape directory, shapes and record dictionary.
Required feature bits 0 to 14 (0x7FFF): NAME_TEXT, DISTINCT_WORDS, SCRIPT_FOLD, WRITTEN_LENGTH, UNPLACED_POSTINGS, UNSIGNED_POSTINGS, COLUMNAR_RECORDS, DELTA_SPATIAL, SHAPES, RECORD_DICTIONARY, ID_FINGERPRINTS, MICRODEGREES, PLACE_DETAILS, APOSTROPHE_WORDS, CJK_PAIRS.
Optional feature bits:
- bit 15 (0x08000) SHORT_PREFIX_LISTS: prefixes of 2 to 6 characters all have posting lists and dictionary blocks are cut at 32 KiB. Clear: prefixes of 2 and 3 characters have head lists only.
- bit 16 (0x10000) COMPACT_DOC_INDEX: the document index keeps an offset every 64 blocks and a 16-bit length per block. Clear: one u64 offset per block.