Module change_point 

Change point detection using the PELT (Pruned Exact Linear Time) algorithm.

This module provides functionality to detect regime shifts in time series data, helping identify when performance characteristics changed significantly.

Algorithm Details

PELT uses a penalty parameter to control sensitivity to change points. The algorithm minimizes: cost + penalty * number_of_change_points, where cost measures how well segments fit the data.

Tuning the Penalty Parameter

• Lower penalty (e.g., 0.5): More sensitive, detects multiple change points
• Higher penalty (e.g., 3.0+): Conservative, only detects major changes
• Default: 0.5 - balanced for detecting multiple significant changes

Configure via .gitperfconfig:

[change_point]
penalty = 0.5  # Global default

[change_point."specific_measurement"]
penalty = 1.0  # Per-measurement override

The scaled penalty used internally is: penalty * log(n) * variance This adapts to data size and variability automatically.

Structs

ChangePoint

A detected change point in the time series

ChangePointConfig

Configuration for change point detection

EpochTransition

Represents an epoch boundary transition

Enums

ChangeDirection

Direction of a detected change

Functions

detect_change_points

Detect change points in a time series using the PELT algorithm.

detect_epoch_transitions

Detect epoch transitions in a sequence of measurements.

enrich_change_points

Convert raw change point indices to enriched ChangePoint structures.