lancedb/docs/web/basics/lance.mdx

---
title: "Lance format"
sidebarTitle: "Lance Format"
description: "Lance is the open format underneath. LanceDB is what implements a lakehouse on top of it."
icon: "/static/assets/logo/lance-logo-gray.svg"
keywords: ["lance", "format", "open standard", "lakehouse", "namespace", "catalog"]
---

Two names, two things. [Lance](https://lance.org/) is an open format and
specification. LanceDB is an implementation built on it.

The distinction matters when you are deciding what you are committing to: the
format is a standard other engines read, and your data stays readable by them
whichever LanceDB you run.

## Lance is the open standard {#standard}

Lance is an open-source, columnar lakehouse format for multimodal AI. It
specifies three things:

| | |
| --- | --- |
| **File format** | how bytes are laid out on disk, optimized for random access without giving up scan throughput |
| **Table format** | how files compose into a versioned table -- schema evolution, deletions, indexes |
| **Namespace specification** | how tables are named, grouped, and resolved -- the catalog layer |

Being a specification rather than a product is the point. Object storage holds
the bytes, and open engines -- PyTorch, Ray, Spark, Trino, DuckDB, Polars --
read them without going through LanceDB at all.

<Card
  title="Lance format documentation"
  icon="/static/assets/logo/lance-logo-gray.svg"
  href="https://lance.org/format"
>
  The format specification: design, layout, and how it enables the multimodal lakehouse.
</Card>

## LanceDB embedded implements the storage half {#embedded}

Connecting to a directory or a bucket gives you a **storage-only lakehouse**.
Your process opens Lance tables directly and does the work itself: it plans and
executes queries, builds indexes, and writes new versions.

```python
db = lancedb.connect("./.lancedb")
db = lancedb.connect("s3://bucket/path")
```

There is no server, which is what makes it embedded -- and also what bounds it.
There is nothing to run work on your behalf, so there are no server-side
[jobs](/basics/jobs), no [SQL endpoint](/basics/apis), and no shared catalog
beyond what the storage layer itself provides.

## LanceDB Enterprise implements the catalog specification {#enterprise}

<Badge color="red">Enterprise</Badge> A `db://` connection talks to a deployment
that implements the **Lance REST Namespace** -- the REST server defined by the
Lance namespace specification, covering the catalog operations in the
[REST API reference](/api-reference/rest).

```python
db = lancedb.connect("db://my-database", api_key="...")
```

Implementing the specification rather than inventing an interface is what makes
the catalog reachable by anything that speaks it, and what makes
[authorization](/governance/authorization), server-side jobs, and
[SQL](/basics/apis) possible: there is now a server to enforce and to run them.

The tables are the same Lance tables either way. What changes is who resolves
their names and who does the work.

## What the format gives you {#capabilities}

These are properties of Lance, so they hold in both forms:

| Capability | What it enables |
| --- | --- |
| Multimodal storage | Images, video, audio, text, embeddings, annotations and metadata in one table. |
| First-class blob API | Large binary objects in blob columns, with lazy reads and streaming byte access. |
| Fast random access and scans | Sample, shuffle, and fetch individual rows without giving up sequential throughput. |
| Zero-copy schema evolution | Add, drop, rename, or alter columns without rewriting existing data files. |
| Versioned tables | Reproduce experiments and tie artifacts to the exact table version they used. |
| Hybrid search and indexing | Vector search, full-text search, and scalar filters over the same dataset. |
| Open interoperability | Object storage underneath, open engines on top. |

Two of them explain behaviour you will meet elsewhere in these docs.

### Versioning is why old data lingers {#versioning}

Every write creates a version, and old versions are not immediately destroyed --
another reader may be part-way through one. A hundred versions are not a hundred
copies of the data, but they do carry a hundred times the metadata, which
eventually slows queries down.

Deletion works the same way: a deleted row is marked in a deletion file rather
than erased, so the data is recoverable for as long as the version is.

[Compaction](/maintenance/compaction) is what resolves both.

### Column evolution is why feature engineering is cheap {#evolution}

Adding a derived column writes only the new column. Existing data -- the
expensive part, the images and videos -- is untouched. That is what makes
[computed columns](/feature-engineering/computed-columns) a reasonable thing to
do repeatedly on a large table rather than a full rewrite each time.

<Card
  title="Learn more about Lance"
  icon="/static/assets/logo/lance-logo-gray.svg"
  href="https://lance.org/quickstart"
>
  Lance is a separate open source project with its own documentation.
</Card>