Amundsen Data Catalog: Lyft's Open Source Tool to Tackle Data Discovery Challenges

June 29th, 2021

What is a data catalog? Image of catalogued books in a library

What is Lyft Amundsen?

Amundsen is a data discovery platform and metadata engine that was developed at Lyft to address the common pain points faced by their data scientists, engineers, and researchers in their typical workflows.

Homegrown by the Lyft engineering team, Amundsen was named after Norwegian explorer Roald Amundsen, who’s most famous for leading the first successful expedition to the South Pole.

Amundsen has improved the productivity of data scientists, analysts, and researchers at Lyft by ~20%

Why did Lyft build Amundsen?

Lyft reported an active rider base of 13.49 million in the first quarter of 2021. Now, imagine this number, in turn, generating a tremendous amount of data to be stored, processed, and analyzed, and also the huge number of people who might be using this data daily to make informed decisions.

At a fundamentally modern data-driven company like Lyft, every interaction is powered by data, and it's impossible to scale sustainably if the data teams are not empowered to productively and effectively use this data.

Lyft recognized this challenge and developed Amundsen, which they introduced in April 2019 as a solution to their data discovery woes.

Amundsen was developed to minimize time spent in discovering and trusting data.

Amundsen was developed to minimize time spent in discovering and trusting data. Source: Lyft Engineering

Lyft created Amundsen to be able to

  1. Capture metadata from all their different data sources
  2. Generate visibility (via metadata) on how data has evolved through its lifecycle
  3. Share this metadata with users by a frontend to enable them to discover, trust and use the data.

Is Amundsen data catalog open source?

Amundsen was open sourced in October 2019, a year following its launching in production at Lyft & is licensed under the Apache License, Version 2.0. A copy of the license can be found here. Here's a roundup of the permissions, limitations and conditions that govern the license.

Amundsen Open Source

Source: Amundsen Github

Amundsen was donated to Linux Foundation AI in July 2020.

How does Amundsen data catalog work?

Amundsen primarily works towards users the ability to discover, trust and understand their data. The various features of Amundsen all work together to achieve the same. Following are Amundsen’s main capabilities:

  • Easy discovery of trusted data
  • Automated & curated metadata
  • Ability to share context with coworkers
  • Learn and understand from data usage

Easy discovery of trusted data

Amundsen helps find data within an organization by a simple text search. The search results go to the extent of showing in-line metadata - that includes description of data and also the last date when it was updated. The page-rank inspired algorithm returns with popularity ranking and also recommendations - highly queried tables are bumped higher for consideration, while least used tables are populated later in the results.

Easy discovery of trusted data

Visual search result with text query. Source: Lyft Engineering

Automated and curated metadata

When a data asset is clicked on, users are shown its detailed description and its behaviour. Detail description includes manually curated information by users. Information about behaviour of the data is generated by grazing through audit logs. By design, users are encouraged to use column level data based on popularity. Users also have visibility on popular users and general profile of the data.

Automated and curated metadata

Metadata describing application context, behaviour and change. Source: Lyft Engineering

Ability to share context with coworkers

One can update descriptions to data assets, thus reducing back and forth between co-workers looking for more context in a particular data asset - for example, updating tables and columns with descriptions may give a data user necessary information about which table is the best fit to a particular query and which particular query in that table is of particular interest.

Ability to share context with coworkers

Manually fed descriptions for better context to viewer. Source: Amundsen Github

Learn and understand from data usage

Users can see which data assets get frequently used, owned, or bookmarked. One can even understand the most common queries for a table by seeing dashboards built on a given table. This is what classifies as behaviour metadata - and is automatically curated.

Learn and understand from data usage

Visibility of relationship between users and resources. Source: Lyft Engineering

The Lyft Amundsen Architecture

Amundsen by Lyft consists of five major components and follows a micro-service architecture:

  1. Metadata Service: Able to handle requests from both frontend service and microservices
  2. Search Service: Backed by elastic search
  3. Frontend Service: Hosts the web application
  4. Databuilder: Ingestion framework which extracts metadata from various sources
  5. Common Components: Library repo which holds common codes among microservices

Metadata Service

Able to handle requests from both frontend service and microservices. Essentially, the metadata is exposed via the front-end service to end users and is also used for other services at Lyft. It is interesting to note that metadata entities are currently modelled as graphs on Amundsen which makes it easier to extend the model when more entities are produced.

Search Service

Backed by elastic search - the search service makes provision for an API to index resources into the search engine, and serve search queries from end users via the front end service. Currently Amundsen supports the following kind of searches:

  • Normal search: Search specifying particular term and resource term
  • Category search: Filtered resources if search term matches a metadata category, relevancy is considered while serving results
  • Wildcard search: Users can do wildcard search over different resources

Frontend Service

Hosts the web application. It consists of two distinct parts:

  • A React Application for composing the user interface
  • A Flask server acting as an intermediary for metadata or search service requests


The ETL Ingestion framework which extracts metadata from various sources. It is said to be highly inspired by Apache Goblin. Each component of the databuilder is highly modularized. Components include:

  • Extractor
  • Transformer
  • Loader
  • Publisher (optional)

Common Components

Library repo which holds common codes for microservices at Amundsen.

Amundsen Architecture

Microservices architecture of Amundsen. Source: Lyft Engineering

Democratizing Data Discovery at Lyft

Amundsen is used by 750 data users at Lyft.

Data teams are diverse. Data engineers, data scientists, analysts, product managers, executives - are all looking for data to process and make informed decisions. True democratization is possible when everyone looking for data resources know exactly what data is available within the system & how they can use it, but that may also pose challenges with respect to data privacy & security. Amundsen seeks to walk the balance between democratization and security by classifying metadata into two groups:

Fundamental metadata

Fundamental metadata like name and description of table and fields, owners, last updated, etc. are visible to all. This enables users to find of its existence and also to understand if it fits their query.

Richer metadata

Richer metadata like column stats, preview, etc. are only available to users with access to data. One can also request access to richer metadata if they are convinced that it's the right fit for them.

Lyft Amundsen Website

Lyft Amundsen GitHub

See Lyft Amundsen Demo

Mail Lists:

Amundsen is widely adopted within Lyft and enjoys a cohesive community of individuals and organizations who have extended and used it. However like any other open source tool - it's made by engineers and for engineers, thus quite technical to set up.

If you are a data consumer or producer and are looking to champion your organization to optimally utilize the value of a modern data stack - while weighing your build vs buy options, it’s worth taking a look at off-the-shelf alternatives like Atlan.

Explore Atlan for free

"It would take six or seven people up to two years to build what Atlan gave us out of the box. We needed a solution on day zero, not in a year or two."

Akash Deep Verma
Akash Deep Verma

Director of Data Engineering

Delhivery: Leading fulfilment platform for digital commerce.

Build vs Buy: Delhivery’s Learnings from Implementing a Data Catalog

Build vs Buy: Delhivery’s Learnings from Implementing a Data Catalog