Skip to content

Latest commit

 

History

History
167 lines (117 loc) · 5.21 KB

README.md

File metadata and controls

167 lines (117 loc) · 5.21 KB
Raw

Python bindings for Lance Data Format

⚠️ Under heavy development

Lance Logo

Lance is a new columnar data format for data science and machine learning

Why you should use Lance

  1. Is order of magnitude faster than parquet for point queries and nested data structures common to DS/ML
  2. Comes with a fast vector index that delivers sub-millisecond nearest neighbors search performance
  3. Is automatically versioned and supports lineage and time-travel for full reproducibility
  4. Integrated with duckdb/pandas/polars already. Easily convert from/to parquet in 2 lines of code

Quick start

Installation

pip install pylance

Make sure you have a recent version of pandas (1.5+), pyarrow (10.0+), and DuckDB (0.7.0+)

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()

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]

*More distance metrics, HNSW, and distributed support is on the roadmap

Python package details

Install from PyPI: pip install pylance # >=0.3.0 is the new rust-based implementation Install from source: maturin develop (under the /python directory) Run unit tests: make test Run integration tests: make integtest

Import via: import lance

The python integration is done via pyo3 + custom python code:

  1. We make wrapper classes in Rust for Dataset/Scanner/RecordBatchReader that's exposed to python.
  2. These are then used by LanceDataset / LanceScanner implementations that extend pyarrow Dataset/Scanner for duckdb compat.
  3. Data is delivered via the Arrow C Data Interface

Motivation

Why do we need a new format for data science and machine learning?

1. Reproducibility is a must-have

Versioning and experimentation support should be built into the dataset instead of requiring multiple tools.
It should also be efficient and not require expensive copying everytime you want to create a new version.
We call this "Zero copy versioning" in Lance. It makes versioning data easy without increasing storage costs.

2. Cloud storage is now the default

Remote object storage is the default now for data science and machine learning and the performance characteristics of cloud are fundamentally different.
Lance format is optimized to be cloud native. Common operations like filter-then-take can be order of magnitude faster using Lance than Parquet, especially for ML data.

3. Vectors must be a first class citizen, not a separate thing

The majority of reasonable scale workflows should not require the added complexity and cost of a specialized database just to compute vector similarity. Lance integrates optimized vector indices into a columnar format so no additional infrastructure is required to get low latency top-K similarity search.

4. Open standards is a requirement

The DS/ML ecosystem is incredibly rich and data must be easily accessible across different languages, tools, and environments. Lance makes Apache Arrow integration its primary interface, which means conversions to/from is 2 lines of code, your code does not need to change after conversion, and nothing is locked-up to force you to pay for vendor compute. We need open-source not fauxpen-source.