Awesome Open Source
Awesome Open Source


This project is a web crawler and search engine for datasets, specifically meant for data augmentation tasks in machine learning. It is able to find datasets in different repositories and index them for later retrieval.

Documentation is available here

It is divided in multiple components:

  • Libraries
    • Geospatial database datamart_geo. This contains data about administrative areas extracted from Wikidata and OpenStreetMap. It lives in its own repository and is used here as a submodule.
    • Profiling library datamart_profiler. This can be installed by clients, will allow the client library to profile datasets locally instead of sending them to the server. It is also used by the apiserver and profiler services.
    • Materialization library datamart_materialize. This is used to materialize dataset from the various sources that Auctus supports. It can be installed by clients, which will allow them to materialize datasets locally instead of using the server as a proxy.
    • Data augmentation library datamart_augmentation. This performs the join or union of two datasets and is used by the apiserver service, but could conceivably be used stand-alone.
    • Core server library datamart_core. This contains common code for services. Only used for the server components. The filesystem locking code is separate as datamart_fslock for performance reasons (has to import fast).
  • Services
    • Discovery services: those are responsible for discovering datasets. Each plugin can talk to a specific repository. Materialization metadata is recorded for each dataset, to allow future retrieval of that dataset.
    • Profiler: this service downloads a discovered dataset and computes additional metadata that can be used for search (for example, dimensions, semantic types, value distributions). Uses the profiling and materialization libraries.
    • Lazo Server: this service is responsible for indexing textual and categorical attributes using Lazo. The code for the server and client is available here.
    • apiserver: this service responds to requests from clients to search for datasets in the index (triggering on-demand query by discovery services that support it), upload new datasets, profile datasets, or perform augmentation. Uses the profiling and materialization libraries. Implements a JSON API using the Tornado web framework.
    • The cache-cleaner: this service makes sure the dataset cache stays under a given size limit by removing least-recently-used datasets when the configured size is reached.
    • The coordinator: this service collects some metrics and offers a maintenance interface for the system administrator.
    • The frontend: this is a React app implementing a user-friendly web interface on top of the API.

Auctus Architecture

Elasticsearch is used as the search index, storing one document per known dataset.

The services exchange messages through RabbitMQ, allowing us to have complex messaging patterns with queueing and retrying semantics, and complex patterns such as the on-demand querying.

AMQP Overview


The system is currently running at You can see the system status at

Local deployment / development setup

To deploy the system locally using docker-compose, follow those step:

Set up environment

Make sure you have checked out the submodule with git submodule init && git submodule update

Make sure you have Git LFS installed and configured (git lfs install)

Copy env.default to .env and update the variables there. You might want to update the password for a production deployment.

Make sure your node is set up for running Elasticsearch. You will probably have to raise the mmap limit.

The API_URL is the URL at which the apiserver containers will be visible to clients. In a production deployment, this is probably a public-facing HTTPS URL. It can be the same URL that the "coordinator" component will be served at if using a reverse proxy (see nginx.conf).

To run scripts locally, you can load the environment variables into your shell by running: . scripts/ (that's dot space scripts...)

Prepare data volumes

Run scripts/ to initialize the data volumes. This will set the correct permissions on the volumes/ subdirectories.

Should you ever want to start from scratch, you can delete volumes/ but make sure to run scripts/ again afterwards to set permissions.

Build the containers

$ docker-compose build --build-arg version=$(git describe) apiserver

Start the base containers

$ docker-compose up -d elasticsearch rabbitmq redis minio lazo

These will take a few seconds to get up and running. Then you can start the other components:

$ docker-compose up -d cache-cleaner coordinator profiler apiserver apilb frontend

You can use the --scale option to start more profiler or apiserver containers, for example:

$ docker-compose up -d --scale profiler=4 --scale apiserver=8 cache-cleaner coordinator profiler apiserver apilb frontend


Import a snapshot of our index (optional)

$ scripts/

This will download an Elasticsearch dump from and import it into your local Elasticsearch container.

Start discovery plugins (optional)

$ docker-compose up -d socrata zenodo

Start metric dashboard (optional)

$ docker-compose up -d elasticsearch_exporter prometheus grafana

Prometheus is configured to automatically find the containers (see prometheus.yml)

A custom RabbitMQ image is used, with added plugins (management and prometheus).

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Python (1,142,969
Dataset (3,633
Search (2,697
Search Engine (1,297
Index (407
Crawling (358
Related Projects