lance-format/lance

Lance is an open lakehouse format for multimodal AI. It contains a file format, table format, and catalog spec that allows you to build a complete lakehouse on top of object storage to power your AI workflows. Lance is perfect for:

1. Building search engines and feature stores with hybrid search capabilities.
2. Large-scale ML training requiring high performance IO and random access.
3. Storing, querying, and managing multimodal data including images, videos, audio, text, and embeddings.

The key features of Lance include:

• Expressive hybrid search: Combine vector similarity search, full-text search (BM25), and SQL analytics on the same dataset with accelerated secondary indices.

• Lightning-fast random access: 100x faster than Parquet or Iceberg for random access without sacrificing scan performance.

• Native multimodal data support: Store images, videos, audio, text, and embeddings in a single unified format with efficient blob encoding and lazy loading.

• Data evolution: Efficiently add columns with backfilled values without full table rewrites, perfect for ML feature engineering.

• Zero-copy versioning: Automatic versioning with ACID transactions, time travel, tags, and branches—no extra infrastructure needed.

• Rich ecosystem integrations: Apache Arrow, Pandas, Polars, DuckDB, Apache Spark, Ray, Trino, Apache Flink, and open catalogs (Apache Polaris, Unity Catalog, Apache Gravitino).

For more details, see the full Lance format specification.

Lance releases frequently because the SDKs, integrations, and performance work are moving quickly. This does not mean the Lance file format changes incompatibly in every release. The Lance file format is identified by the data_storage_version stored in each dataset, and stable storage versions are a long-term compatibility contract.

• Once a dataset is written with a stable data_storage_version, future Lance releases will continue to support reading that storage version.
• SDK and API compatibility is separate from file format compatibility. SDK/API changes follow semantic versioning and are documented in the migration guide.
• Older Lance releases may not understand file format versions introduced later. If you run mixed Lance versions, pin data_storage_version for deterministic writes.
• The next file format alias is unstable and should only be used for experimentation, never for production data.

For production, write data with a stable data_storage_version. See the format versioning guide for the current compatibility matrix.

Installation

pip install pylance

To install a preview release:

pip install --pre --extra-index-url https://pypi.fury.io/lance-format pylance

Converting to Lance

import lance

import pandas as pd
import pyarrow as pa
import pyarrow.dataset

df = pd.DataFrame({"a": [5], "b": [10]})
uri = "/tmp/test.parquet"
tbl = pa.Table.from_pandas(df)
pa.dataset.write_dataset(tbl, uri, format='parquet')

parquet = pa.dataset.dataset(uri, format='parquet')
lance.write_dataset(parquet, "/tmp/test.lance")

Reading Lance data

dataset = lance.dataset("/tmp/test.lance")
assert isinstance(dataset, pa.dataset.Dataset)

Pandas

df = dataset.to_table().to_pandas()
df

DuckDB

import duckdb

# If this segfaults, make sure you have duckdb v0.7+ installed
duckdb.query("SELECT * FROM dataset LIMIT 10").to_df()

Vector search

Download the sift1m subset

wget ftp://ftp.irisa.fr/local/texmex/corpus/sift.tar.gz
tar -xzf sift.tar.gz

Convert it to Lance

import lance
from lance.vector import vec_to_table
import numpy as np
import struct

nvecs = 1000000
ndims = 128
with open("sift/sift_base.fvecs", mode="rb") as fobj:
    buf = fobj.read()
    data = np.array(struct.unpack("<128000000f", buf[4 : 4 + 4 * nvecs * ndims])).reshape((nvecs, ndims))
    dd = dict(zip(range(nvecs), data))

table = vec_to_table(dd)
uri = "vec_data.lance"
sift1m = lance.write_dataset(table, uri, max_rows_per_group=8192, max_rows_per_file=1024*1024)

Build the index

sift1m.create_index("vector",
                    index_type="IVF_PQ",
                    num_partitions=256,  # IVF
                    num_sub_vectors=16)  # PQ

Search the dataset

# Get top 10 similar vectors
import duckdb

dataset = lance.dataset(uri)

# Sample 100 query vectors. If this segfaults, make sure you have duckdb v0.7+ installed
sample = duckdb.query("SELECT vector FROM dataset USING SAMPLE 100").to_df()
query_vectors = np.array([np.array(x) for x in sample.vector])

# Get nearest neighbors for all of them
rs = [dataset.to_table(nearest={"column": "vector", "k": 10, "q": q})
      for q in query_vectors]

We used the SIFT dataset to benchmark our results with 1M vectors of 128D

1. For 100 randomly sampled query vectors, we get <1ms average response time (on a 2023 m2 MacBook Air)

1. ANNs are always a trade-off between recall and performance

We create a Lance dataset using the Oxford Pet dataset to do some preliminary performance testing of Lance as compared to Parquet and raw image/XMLs. For analytics queries, Lance is 50-100x better than reading the raw metadata. For batched random access, Lance is 100x better than both parquet and raw files.

The machine learning development cycle involves multiple stages:

graph LR
    A[Collection] --> B[Exploration];
    B --> C[Analytics];
    C --> D[Feature Engineer];
    D --> E[Training];
    E --> F[Evaluation];
    F --> C;
    E --> G[Deployment];
    G --> H[Monitoring];
    H --> A;

Traditional lakehouse formats were designed for SQL analytics and struggle with AI/ML workloads that require:

• Vector search for similarity and semantic retrieval
• Fast random access for sampling and interactive exploration
• Multimodal data storage (images, videos, audio alongside embeddings)
• Data evolution for feature engineering without full table rewrites
• Hybrid search combining vectors, full-text, and SQL predicates

While existing formats (Parquet, Iceberg, Delta Lake) excel at SQL analytics, they require additional specialized systems for AI capabilities. Lance brings these AI-first features directly into the lakehouse format.

A comparison of different formats across ML development stages: