Skip to main content


The Datalayer is the principle point of entry in superduperdb for:

  • Communicating with the database
  • Instructing models and other components to work together with the database
  • Accessing and storing meta-data about your superduperdb models and data

Technically, the Datalayer "wires together" several important backends involved in the AI workflow:

  • Querying the database via the databackend
  • Storing and retrieving serialized model-weights and other artifacts from the artifact store
  • Storing and retrieval important meta-data, from the meta-data store and information about models and other components which are to be installed with superduperdb
  • Performing computations over the data in the databackend using the models saved in the artifact store
from superduperdb import superduper

db = superduper()

# <superduperdb.backends.mongodb.data_backend.MongoDataBackend at 0x1562815d0>

# <superduperdb.backends.mongodb.artifacts.MongoArtifactStore at 0x156869f50>

# <superduperdb.backends.mongodb.metadata.MongoMetaDataStore at 0x156866a10>

# <superduperdb.backends.local.LocalComputeBackend 0x152866a10>

Our aim is to make it easy to set-up each aspect of the Datalayer with your preferred connections/ engines.


The databackend typically connects to your database (although superduperdb also supports other databackends such as a directory of pandas dataframes), and dispatches queries written in an query API which is compatible with that databackend, but which also includes additional aspects specific to superduperdb.

Read more here.

The databackend is configured by setting the URI CFG.databackend in the configuration system.

We support the same databackends as supported by the ibis project:

Artifact Store

The artifact-store is the place where large pieces of data associated with your AI models are saved. Users have the possibility to configure either a local filesystem, or an artifact store on MongoDB gridfs:

For example:

CFG.artifact_store = 'mongodb://localhost:27017/documents'


CFG.artifact_store = 'filesystem://./data'

Metadata Store

The meta-data store is the place where important information associated with models and related components are kept:

  • Where are the data artifacts saved for a component?
  • Important parameters necessary for using a component
  • Important parameters which were used to create a component (e.g. in training or otherwise)

Similarly to the databackend and artifact store, the metadata store is configurable:

CFG.metadata = 'mongodb://localhost:27017/documents'

We support metadata store via:

  1. MongoDB
  2. All databases supported by SQLAlchemy. For example, these databases supported by the databackend are also supported by the metadata store.

Compute backend

The compute-backend is designed to be a configurable engine for performing computations with models. We support 2 backends:

  • Local (default: run compute in process on the local machine)
  • dask (run compute on a configured dask cluster)

Default settings

In such cases, the default configuration is to use the same configuration as used in the databackend.

I.e., for MongoDB the following are equivalent:

db = superduper('mongodb://localhost:27018/documents')


db = superduper(

Whenever a database is supported by the artifact store and metadata store, the same behaviour holds. However, since there is no general pattern for storing large files in SQL databases, the fallback artifact store is on the local filesystem. So the following are equivalent:

db = superduper('sqlite://<my-database>.db')


from superduperdb.backends.local.compute import LocalComputeBackend

db = superduper(

Key methods

Here are the key methods which you'll use again and again:


This method executes a query. For an overview of how this works see here.


This method adds Component instances to the db.artifact_store connection, and registers meta-data about those instances in the db.metadata_store.

In addition, each sub-class of Component has certain "set-up" tasks, such as inference, additional configurations, or training, and these are scheduled by db.add.

This methods displays which Component instances are registered with the system.


This method removes a Component instance from the system.

Additional methods


Validate your components (mostly models)


Infer predictions from models hosted by superduperdb. Read more about this and about models here.