Basic Objects

Model

In drifting, 'Model' name is used for predictive algorithm used by a user outside this package. Model can be understood as the algorithm, its weights (parameters), inference graph.

Drift Detection is a process of measuring the Model or data drift. Usually, Drift Detector is fitted for each training of a given Model

Drift Detection Server (DDS)

Drift Detection Server is an API-first Python server implemented with FastAPI. DDS serves one or more Drift Detectors. It also exposes the endpoints for fitting the Drift Detector, getting metadata, and managing all the Projects.

Drift Detectors

Drift Detector is an algorithm that, based on reference data, estimates if the statistical difference of newly arrived data is significant. Drift Detectors use Alibi Detect package for the statistical calculations.

Project

Project consists of 3 elements:

1. detector_name - unique name that allows to distinguish the model and drift detector from others
2. drift_type - the type of a drift (image, text, tabular)
3. Configuration - configuration that allows the Drift Detector to be loaded.

Drift detector for a detector_name should be trained once and is immutable.