lance-format

Lance v9 reference

Lance is an open columnar format for multimodal AI - "a columnar data format that is 100x faster than Parquet for random access." It is not one format but a stack of interoperating specs: a file format, a table format, index formats, catalog specs, and a namespace client spec. The Rust workspace at lance-format/lance implements all of them plus Python (pylance) and Java bindings.

This skill tracks v9.0.0-beta.10 (the lance-format/lance git tag), the current development frontier. Pin against tags, not main - Lance ships beta tags every few days and next-format encodings can change. v8.0.0 is the concurrent stabilizing release (v8.0.0-rc.3, 2026-06-30; no final tag yet at time of writing) - if you need a stable pin rather than the v9 dev betas, track the v8.0.0 rc/final line, whose format and API are the frozen predecessor of what this reference describes.

The deep reference is references/lance-reference.md. Load it for any concrete schema, parameter, proto, or constraint. This file is the orientation: read it first, then jump into the reference section you need.

Lance vs LanceDB

These are two different things and conflating them produces wrong answers.

• Lance - the format and engine. The lance-format/lance repo; the lance /lance-* Rust crates; pylance. It gives you datasets, the file/table format, indexes, commits, scans. Consumed directly by DuckDB, Polars, Ray, Spark, PyTorch, DataFusion, or your own Rust/Python code. This skill is about Lance.
• LanceDB - a separate database product (lancedb/lancedb) built on top of Lance. It adds a query-builder API, an embedding registry, rerankers-as-API, multi-language SDK parity, and managed Cloud / Enterprise tiers. Not covered here.

If you are linking the lance crate in Cargo.toml, you are using Lance directly - use this skill. If a question is about LanceDB internals, the storage layer underneath it is still Lance, so this skill remains the authority for the format itself.

The crate workspace

25 crate directories under rust/. lance is the public entry point; the rest are layers beneath it. Full table with descriptions and citations in references/lance-reference.md section 2.

| Crate | Role | |-------|------| | lance | Public entry point - Dataset, scanner, indexes, commits | | lance-table | Table format - manifest, feature flags, commit handlers, row IDs | | lance-file | File format - file reader/writer | | lance-encoding | Structural encodings, compression (internal, not for external use) | | lance-index | Scalar / vector / FTS / system indexes | | lance-io | Object store, I/O schedulers | | lance-core | Shared Error/Result, cache, datatypes | | lance-datafusion | DataFusion glue (exec, expr, planner, UDFs) | | lance-linalg | SIMD L2 / dot / cosine / hamming kernels | | lance-select | Row-selection primitives - RowAddrMask, RowIdMask, IndexExprResult (extracted from lance-core/lance-index in v7.1.0-beta.2) | | lance-tokenizer | FTS tokenizer stack (simple, ngram, jieba, lindera, stemmers) | | lance-derive | #[derive(DeepSizeOf)] proc-macro for Arrow-aware memory accounting (new in v8, PR #6229; replaced the external deepsize crate) | | lance-geo | Geospatial UDFs (feature-gated geo) | | lance-namespace / -impls / -datafusion | Namespace trait, Directory/REST impls, DataFusion catalog bridge | | lance-arrow, lance-tools, fsst, lance-bitpacking, ... | Arrow extensions, CLI, compression sub-crates |

All share version = "9.0.0-beta.10" except lance-arrow-scalar, which is pinned at 58.0.0 to track Arrow. Workspace: edition 2024, rust-version = 1.91.0, resolver = "3"; notable deps arrow 58, datafusion 53, opendal 0.57, jieba-rs 0.10, lance-namespace-reqwest-client 0.8.6, itertools 0.14. Python bindings now require Python 3.10+ (3.9 dropped in v9, PR #7345).

File format versions

The file format carries a single major.minor version. Selected per-dataset at creation via data_storage_version and fixed once the dataset exists (to change it, rewrite the dataset).

| Version | Status | Notes | |---------|--------|-------| | 0.1 (legacy) | read-only | Original format; no longer writable | | 2.0 | stable | Removed row groups; null support for lists/FSL/primitives | | 2.1 | current default (stable) | Adaptive structural encodings; better integer/string compression; nulls in struct fields; better nested random access. Default since Lance 5.0.0 | | 2.2 | unstable | Map type, Blob v2, VariablePackedStruct, larger mini-blocks. Required for Map and Blob v2; the real experimental frontier - encodings may still change | | 2.3 | unstable (next) | The current next alias target (V2_3 in the enum). Scaffolding only - no distinct encodings yet (6 refs vs 98 for 2.2 in lance-encoding); the docs version table still lists only 2.2 |

stable resolves to the default (2.1); next now resolves to 2.3 (not 2.2) in the running Lance release - pin an explicit number for deterministic behavior. In the version ladder 2.2 sits below next, so the code does not flag 2.2 as unstable. As of v9 the docs version table (docs/src/format/file/versioning.md) lists 2.3 as the unstable row and no longer labels 2.2 unstable (it now reads "2.2-era storage features") - code and docs finally agree that 2.3 is the unstable frontier, though 2.2 still carries the concrete experimental encodings (Map, Blob v2, VariablePackedStruct).

What's new in v9

The v8 -> v9 boundary is a light major bump: structurally v9 is nearly identical to v8 (same 25 crates, 15 transaction ops, file-format enum with next => 2.3 and default 2.1, CommitConfig.num_retries still 20, arrow 58 / datafusion 53 / opendal 0.57 / jieba 0.10 unchanged). The major version was auto-triggered by Lance's breaking-change-label detector (ci/check_breaking_changes.py), fired by two PRs: Python 3.9 was dropped (minimum now 3.10, #7345) and alter_columns now fails fast when you cast a column that has an index attached - you must drop_index() first instead of relying on the old silent drop/invalidate (#7158). A third breaking change rode the already-bumped series: the FM-Index proto message was renamed FMIndexIndexDetails -> FMIndexDetails (#7397), which makes existing FM indexes unreadable. One public Rust-API removal: as_vector_index is gone from the Index trait (#7392) - downcast via as_any().

Net-new in v9: a hamming clustering utility for near-duplicate detection (SIMD union-find over 64-bit binary hashes, #7379); **COUNT() pushdown now works on stable-row-id datasets (#7360); per-column blob size thresholds (lance-encoding:blob-inline-size-threshold / ...-dedicated-size-threshold, #7269); tunable 32k miniblock chunks via LANCE_MINIBLOCK_MAX_VALUES (#7356, default still 4096); an icu/split FTS tokenizer variant (#7474); distributed LabelList index builds (#7223); the ngram index now accelerates regex and infix LIKE (#7139); and cleanup-explain plus fragment-reuse remap are now exposed to Python and Java* (#7248, #7438). Full delta in references/lance-reference.md section 14.

The v7 -> v8 boundary (the predecessor line) unified all index builds onto one segment-based lifecycle: bitmap migrated to the segment workflow (#6869), the standalone IndexSegmentBuilder API was removed (#6997), distributed BTree moved to the segmented framework (#7013), file writers' finish() began returning FileWriteSummary (#7096), and describe_indices()/list_indices() were reworked (#6903, #7129). v8 also added the lance-derive crate (#6229), the FM-Index scalar index, multi-bit IVF_RQ (num_bits 1..=9), the public vector-search approx_mode (fast/normal/accurate), and the Volcengine TOS (tos://) and feature-gated GooseFS (goosefs://) object stores. The v7 era - MemWAL, branches, the geo/RTree index, the lance-select crate, ICU FTS - all carries forward.

Navigating the reference

references/lance-reference.md is the full v9 reference, regrounded against the v9.0.0-beta.10 source. Load the section for your task:

1. What Lance is - the lakehouse spec stack
2. Crate workspace - all 25 crates, what each does, the public entry point
3. File format - versions, container layout, structural encoding (mini-block / full-zip / constant / blob page types), compression schemes, blob encoding
4. Data types - Arrow type coverage, FixedSizeList for vectors, JSON (JSONB), blob, ML extension arrays (bfloat16, image types)
5. Table format - dataset directory layout, manifest contents, fragments, deletion files, base paths
6. Schema evolution - field IDs, zero-copy column add/drop/alter, why old rows read NULL
7. Versioning, tags, branches - manifest versions, time travel, tag pinning, branches
8. Row IDs - row address vs stable row ID, lineage, change-data-feed columns
9. Transactions and concurrency - the 15 transaction ops, OCC retry/rebase, commit handlers (conditional-put, DynamoDB), conflict resolution matrix
10. MemWAL - shards, MemTable/WAL/flush, the appender/tailer/flusher model, fencing
11. Indexes - vector (IVF/HNSW/PQ/SQ/RQ, multi-bit RQ), scalar (btree/bitmap/bloom/ labellist/ngram/zonemap/FM-Index), full-text (BM25, tokenizers), geo/RTree
12. Distributed write and indexing - two-phase commits, segment-based index builds
13. Object store - URI schemes, storage options, per-backend config
14. What changed - the full v7 -> v8 -> v9 delta
15. Capability matrix - what Lance can and cannot do
16. Source map - where each spec and proto lives in the repo

Maintenance

Citations in references/lance-reference.md are path:line relative to the lance-format/lance repo; build a permalink as https://github.com/lance-format/lance/blob/v9.0.0-beta.10/<path>.

To refresh: git -C ~/pjv/lance-format/lance fetch --tags, check out the newest v9* tag (or the v8.0.0 rc/final line for a stable pin), re-read the format spec under docs/src/format/ and the user guide under docs/src/guide/, re-verify the crate workspace, and bump metadata.upstream plus every v9.0.0-beta.10 reference. Line numbers in citations drift between tags - treat them as approximate.