@datagrok/seldon v0.0.4

Seldon

Seldon is a package for the Datagrok platform. It allows connecting to a Seldon Cluster with deployed machine learning models and easily apply them within Datagrok.

Overview

Seldon is a popular ML Ops platform for managing and using containerized machine learning models. Many of such models are of great use directly inside Datagrok through its built-in custom machine learning models infrastructure. The package Seldon uses this infrastructure to add support for running arbitrary Seldon models simply by pointing to the cluster of interest. Models with their inputs and outputs need not be known in advance in order to be used. It is also possible to enumerate and connect to Seldon models programmatically in your Datagrok extensions and applications in TypeScript/JavaScript.

Since version 1.3 Seldon supports metadata, or prediction schema, which allows understanding, for any given model, its inputs, outputs, their order, types, inner structure (one-hot, probabilistic) and other traits. Seldon package supports any Seldon-deployed model which provides for this metadata.

Models are deployed to a Seldon Cluster into namespaces and are called deployments. In Seldon there is a new Seldon Deployment SDK which provides for all the relevant enumeration of these namespaces and deployments (models).

Currently, the package support identification with Keycloak.

Usage

After deploying the package Seldon, set up the followng credentials for it (the values are provided as examples):

• seldonUser
• seldonPassword
• seldonHost: https://models.YOUR-SERVER.com/seldon-deploy/api/v1alpha1
• seldonOIDCServer: https://isaac-keycloak.YOUR-SERVER.com/auth/realms/orgName
• seldonClientID: seldon-api
• seldonNamespace: test-namespace

Roadmap

Alpha version

Is built for pre-v1.3 deployments which don't have model metadata provided. The demo mocks the metadata for a popular sklearn predictive model iris.

Open iris.csv in Datagrok, activate the top menu and find the item Seldon | Apply. The list of deployments (models) will be loaded for a pre-specified namespace. Select the one corresponding to iris and hit Ok. The iris dataframe will be shortly filled in with predictions coming from the Seldon model.

Beta version

• Supports enumeration of the namespaces
• Supports arbitrary metadata parsing for Seldon versions >= 1.3

Release 1.0

• Supports all kinds of data types supported by Seldon metadata