Continual is a fully-managed cloud platform designed to simplify operational AI within the enterprise. This page provides an overview of its architecture, components, and features.
Relationship to your data warehouse¶
Continual has a hybrid architecture where your data — your features and predictions — live in your existing data warehouse where they are governed by your existing data policies. Continual access this data temporarily for training and prediction purposes but does not otherwise replicate or store the data.
Continual consistents of a control plane responsible for collaboration, metadata storage, and orchestration, and a data plane responsible for model training and prediction.
Continual generally attempts to push as much computation into the underlying data platform for increased performance and security. For instance, features you define in SQL are executed inside your underlying data platform. Increased support for push down of training and inference is under development on platforms that support it, such as Snowflake with Snowpark and Databricks.
Continual has three core components that together provide an end-to-end solution for operational AI on the modern data stack.
Continual Feature Store - The Continual Feature Store allows users to define and collaborate on the feature sets leveraging all the data in their data warehouse and existing dbt models, if desired. Users can gather features into shared business entities and then link them to models to avoid duplication of work or costly errors. Continual handles joining data together in a time-aware fashion to avoid feature and target leakage. With a shared feature store, users can quickly develop new models and improve existing ones.
Continual AI Engine - Continual's automated AI engine enables users of all skill levels to rapidly build state-of-the-art predictive models without complex engineering. Models are automatically updated when new features are added to Continual and Continual maintains the lifecycle of both the model and predictions over time.
Continual MLOps and XAI - Continual's MLOps and XAI capabilities allow you to monitor your models and predictions over time. You can understand why your models are making the predictions they are, how performance changes over time, and when data or models have drifted enough to require retraining. See how changes to models or features affect performance, all in a GitOps friendly workflow that makes it simple to build new use cases or update existing ones.
The diagram below illustrates how data typically flows through Continual to generate predictions that power downstream applications.
Source Tables represent the data models you have created for other purposes. Features represent transformations of these source tables into features that are useful for predictive modeling. Predictions are transformations of features into predictions. Finally, Analytic Tables represent downstream processing of these predictions, or another data, for use by applications. Continual provides a declarative workflow to manage features and define models which generate predictions directly on your existing data warehouse.