lamindb 0.46.1

LaminDB: Manage R&D data & analyses.

LaminDB

Open-source data lake & feature store for biology.

Manage your existing R&D data & analyses in your existing infrastructure.

Public beta: Currently only recommended for collaborators as we still make breaking changes.

Update 2023-06-14:

- We completed a major migration from SQLAlchemy/SQLModel to Django, available in 0.42.0.
- The last version before the migration is 0.41.2.

Introduction

LaminDB is a free & open-source Python library allowing you to:

• Manage files & datasets while tracking provenance across pipelines, notebooks & apps.
• Manage biological registries, ontologies, features & schemas.
• Build in data integrity through automated validation and idempotent & ACID operations.
• Use one API for metadata & data to query, search & look-up and save, load & stream data.
• Collaborate across a mesh of LaminDB instances (share them in a hub).

LaminApp is a data management app built on LaminDB, deployable in your infrastructure. Think LaminApp ~ GitHub and LaminDB ~ git.

LaminApp, support & code templates for a BioTech data & analytics platform are currently only available on an enterprise plan.

Usage overview & quickstart

If you'd like to run the following snippets: the setup takes 2 min.

Initialize a data lake instance with local or cloud default storage on the CLI:

$ lamin init --storage ./mydata   # or s3://my-bucket, gs://my-bucket, etc.

Import lamindb:

import lamindb as ln

Store, query, search & load data objects

Store a DataFrame in default storage:

df = pd.DataFrame({"feat1": [1, 2], "feat2": [3, 4]})  # AnnData works, too

ln.File(df, name="My dataset1").save()  # create a File object and save/upload it

You have the full power of SQL to query for metadata, but the simplest query for a file is:

file = ln.File.select(name="My dataset1").one()  # get exactly one result

If you don't have specific metadata in mind, run a search:

ln.File.search("dataset1")

Once you queried or searched it, load a file back into memory:

df = file.load()

Or get a backed accessor to stream its content from the cloud:

backed = file.backed()  # currently works for AnnData, zarr, HDF5, not yet for DataFrame

Store, query & search files

The same API works for any file:

file = ln.File("s3://my-bucket/images/image001.jpg")  # or a local path
file.save()  # register the file

Query by key (the relative path within your storage):

file.select(key_startswith="images/").df()  # all files in folder "images/" in default storage

Auto-complete categoricals

When you're unsure about spellings, use a lookup object:

users = ln.User.lookup()
ln.File.select(created_by=users.lizlemon)

Track & query data lineage

In addition to basic provenance information (created_by, created_at, created_by), you can track which notebooks, pipelines & apps transformed files.

Notebooks

Track a Jupyter Notebook:

ln.track()  # auto-detect & save notebook metadata
ln.File("my_artifact.parquet").save()  # this file is now aware that it was saved in this notebook

When you query the file, later on, you'll know from which notebook it came:

file = ln.File.select(name="my_artifact.parquet").one()  # query for a file
file.transform  # the notebook with id, title, filename, version, etc.
file.run  # the specific run of the notebook that created the file

Alternatively, you can query for notebooks and find the files written by them:

transforms = ln.Transform.select(  # all notebooks with 'T cell' in the title created in 2022
    name__contains="T cell", type="notebook", created_at__year=2022
).all()
ln.File.select(transform__in=transforms).df()  # the files created by these notebooks

Pipelines

This works like for notebooks just that you need to provide pipeline metadata yourself.

To save a pipeline to the Transform registry, call

ln.Transform(name="Awesom-O", version="0.41.2").save()  # save a pipeline, optionally with metadata

Track a pipeline run:

transform = ln.Transform.select(name="Awesom-O", version="0.41.2").one()  # select pipeline from the registry
ln.track(transform)  # create a new global run context
ln.File("s3://my_samples01/my_artifact.fastq.gz").save()  # file gets auto-linked against run & transform

Now, you can query for the latest pipeline runs:

ln.Run.select(transform=transform).order_by("-created_at").df()  # get the latest pipeline runs

Run inputs

To track run inputs, pass is_run_input to any File accessor: .stage(), .load() or .backed(). For instance,

file.load(is_run_input=True)

You can also track inputs by default by setting ln.settings.track_run_inputs = True.

Load your data lake from anywhere

If provided with access, others can load your data lake via a single line:

$ lamin load myaccount/myartifacts

Manage biological registries

lamin init --storage ./bioartifacts --schema bionty

...

Track biological features

...

Track biological samples

...

Manage custom schemas

1. Create a GitHub repository with Django ORMs similar to github.com/laminlabs/lnschema-lamin1
2. Create & deploy migrations via lamin migrate create and lamin migrate deploy

It's fastest if we do this for you based on our templates within an enterprise plan, but you can fully manage the process yourself.

Setup

Installation

pip install lamindb  # basic data management

You can configure the installation using extras, e.g.,

pip install lamindb[jupyter,bionty,fcs,aws]

Supported extras are:

jupyter  # Track Jupyter notebooks
bionty   # Manage basic biological entities
fcs      # Manage .fcs files (flow cytometry)
zarr     # Store & stream arrays with zarr
aws      # AWS (s3fs, etc.)
gcp      # Google Cloud (gcfs, etc.)

Sign up

Why do I have to sign up?

• Data lineage requires a user identity (who modified which data when?).
• Collaboration requires a user identity (who shares this with me?).

Signing up takes 1 min.

We do not store any of your data, but only basic metadata about you (email address, etc.) & your LaminDB instances (S3 bucket names, etc.).

• Sign up: lamin signup <email>
• Log in: lamin login <handle>

How does it work?

Dependencies

LaminDB builds semantics of R&D and biology onto well-established tools:

• SQLite & Postgres for SQL databases using Django ORM (previously: SQLModel)
• S3, GCP & local storage for object storage using fsspec
• Configurable storage formats: pyarrow, anndata, zarr, etc.
• Biological knowledge sources & ontologies: see Bionty

LaminDB is open source.

Architecture

LaminDB consists of the lamindb Python package (repository here) with its components:

• bionty: Basic biological entities (usable standalone).
• lamindb-setup: Setup & configure LaminDB, client for Lamin Hub.
• lnschema-core: Core schema, ORMs to model data objects & data lineage.
• lnschema-bionty: Bionty schema, ORMs that are coupled to Bionty's entities.
• lnschema-lamin1: Exemplary configured schema to track samples, treatments, etc.
• nbproject: Parse metadata from Jupyter notebooks.

LaminHub & LaminApp are not open-sourced, and neither are templates that model lab operations.

Lamin's packages build on the infrastructure listed above.

Notebooks

• Find all guide notebooks here.
• You can run these notebooks in hosted versions of JupyterLab, e.g., Saturn Cloud, Google Vertex AI, Google Colab, and others.
• Jupyter Lab & Notebook offer a fully interactive experience, VS Code & others require using the CLI to track notebooks: lamin track my-notebook.ipynb

Documentation

Read the docs.