Concepts

Hydrosphere platform can be summarized in the following diagram. It’s a set of microservices that work together to let you manage machine learning models reliably and at scale.

System

Services

Manager

Manager is a service component, responsible for model cataloging, building, provisioning servables and applications, working with models metadata and basically handling all the resources.

Docker Registry

Each model is represented as a Docker image. Those images are stored in the configured Docker registry. Hydrosphere can use the default registry bundled with the Hydrosphere installation or use external one.

PostgreSQL

Hydrosphere stores information about models, applications, runtimes inside PostgreSQL database.

Gateway

Gateway is a service component responsible for handling prediction requests and routing them among model services. Gateway maps model endpoint name to the corresponding container. Whenever it receives a request it communicates with that container by gRPC protocol.

UI

UI is a service component responsible for showing off user-friendly interfaces of models, deployed applications as well as monitoring charts and profiles.

Sonar

Sonar is a service component responsible for monitoring your models during inference time. It allows you to evaluate how your model behaves under production load, i.e. is there a concept drift occurred in the production data (so your model needs to be retrained); how many outliers are there in the production data; how distribution of your training data is compared with the distribution of the production data?

To learn more about how Sonar works, refer to this page.

Note

Currently is not available in public distribution.

Abstractions

During the interaction with the platform you will encounter with the following entities: Models and Applications.

Models

Model is a machine learning model or a processing function that consume provided inputs and produce predictions/transformations. To learn more about models visit this page.

Applications

Application is a publicly available HTTP/gRPC/Kafka endpoint to reach your models. To learn more about applications visit this page.