Glossary¶
- VCS
- Version control system. A service that stores model source code for development and deployment procedures (e.g. a GitHub Repository).
- Trained Model Binary
- An archive containing a trained ML/AI model (inference code, model weights, etc). Odahu defines a format for these binaries. See <ref_model_format.html>
- Trainer
- Application that uses model source code, Data Bindings, Connections and Training Hyperparameters to produce a Trained Model Binary.
- Data Binding
- Reference to remote data (e.g. files from S3) should be placed for a Train process.
- Connection
- Credentials for an external system. For example: Docker Registry, cloud storage location, etc.
- Training Hyperparameters
- Parameter for Training process. For example, count of epochs in evolution algorithms.
- Train
- A containerized process that converts model source code, Data Bindings, Connections and Training Hyperparameters to Trained Model Binary using a Trainer defined in a Trainer Extension
- Trainer Extension
- A pluggable Train implementation.
- Packager
- Containerized application that uses a Trained Model Binary and Connections and converts them into a target Archive. Typically this is a Docker image with REST API.
- Package
- Containerized process which turns a Trained Model Binary into a Docker image with REST API using a Packager Extension.
- Packager Extension
- A pluggable Package implementation.
- Deployer
- Containerized application that uses the results of a Package process and Connections to deploy a packaged model on a Kubernetes cluster.
- Deploy
- Containerized process that deploys results of a Package operation to Kubernetes cluster with a REST web service.
- Trainer Metrics
- Metrics set by Trainer code during Train (e.g. accuracy of model). These metrics can be used for querying and comparing Train events.
- Key/value value pairs that are set by Trainer code (e.g. type of algorithm). Can be used for querying and comparing Train runs.
- General Python Prediction Interface
- Format of storing models, written in a Python language
- MLflow Trainer
- Integration of MLflow library for training models, written in a Python. Details - MLFlow Trainer
- REST API Packager
- Integration for packing trained models into Docker Image with served via REST API
- API service
- API for managing Odahu Platform resources for cloud deployed Platform
- Operator
- A Kubernetes Operator that manages Kubernetes resources (Pods, Services and etc.) for Odahu Train, Package, and Deploy instances.
- Prediction
- A deployed model output, given input parameters.
- Model prediction API
- API provided by deployed models to allow users to request predictions through a web service.
- Prediction Feedback
- Feedback versus the previous prediction, e.g. prediction correctness.
- Model Feedback API
- An API for gathering Prediction Feedback
- Feedback aggregator
- A service that provides a Model Feedback API and gathers input and output prediction requests
- Odahu-flow SDK
- An extensible Python client library for API service, written in Python language. Can be installed from PyPi.
- Odahu-flow CLI
- Command Line Interface for API service, written in Python. Can be installed from PyPi. It uses the Odahu-flow SDK.
- Plugin for JupyterLab
- A odahu-specific plugin that provides Odahu Platform management controls in JupyterLab.
- Plugin for Jenkins
- A library for managing Odahu Platform resources from Jenkins Pipelines.
- Plugin for Airflow
- A library that provides Hooks and Operators for managing Odahu Platform resources from Airflow.
- Model Deployment Access Role Name
- Name of scope or role for accessing model deployments.
- JWT Token
- A JSON Web Token that allows users to query deployed models and to provide feedback. This token contains an encoded role name.
- A/B testing
- Process of splitting predictions between multiple Model Deployments in order to compare prediction metrics and Model Feedback for models, which can vary by source code, dataset and/or training hyperparameters
- Odahu distribution
- A collection of Docker Images, Python packages, or NPM packages, which are publicly available for installation as a composable Odahu Platform.
- Odahu Helm Chart
- A YAML definition for Helm that defines a Odahu Platform deployed on a Kubernetes cluster.
- Odahu-flow’s CRDs
Objects that API service creates for actions that require computing resources to be stored. For example: connections, Trains, etc.
These objects are Kubernetes Custom Resources and are managed by operator.
- Identity Provider (idP)
- A component that provides information about an entity (user or service).
- Policy Enforcement Point (PEP)
- A component that enforces security policies against each request to API or other protected resources.
- Policy Decision Point (PDP)
- A component that decides whether the request (action in the system) should be permitted or not.