moon-change-point

moon-change-point is a MoonBit library for online and offline statistical change-point detection in production telemetry and time-series pipelines. It targets WASM, WASM-GC, JavaScript, and native builds through the standard MoonBit toolchain.

Project positioning

The library provides composable detection primitives for mean shifts, variance changes, spikes, trends, distribution drift, and multivariate changes. It also includes the surrounding data-quality, streaming, evaluation, replay, and alerting utilities needed to move from a detector score to an operational signal.

Core capabilities

• Online detection: CUSUM, Page-Hinkley, simplified Bayesian online detection, EWMA, Robust-Z, IQR, trend, variance, seasonal, rank, and distribution detectors.
• Streaming pipelines: bounded windows, time aggregation, watermark-based late-data handling, multi-metric monitoring, suppression, ensembles, and multi-scale evidence.
• Offline analysis: binary segmentation, dynamic programming, split inspection, change-point metrics, tolerance-aware evaluation, and piecewise-constant error.
• Multivariate monitoring: vector statistics, projection detection, Mahalanobis distance, covariance/correlation analysis, and correlated-change scoring.
• Operational utilities: data-quality contracts, threshold calibration, bootstrap estimates, deterministic signal generation, replay comparison, explanations, SLO tracking, incident clustering, alert routing, rollout guardrails, and canary plans.
• Integration outputs: stable JSON Lines, CSV, Prometheus, Markdown, and summary JSON exports for logs, notebooks, CI artifacts, and dashboards.

Quick start

Add the package to a MoonBit module:

moon add Zy789kl/moon-change-point

Use a detector through the public package API:

///|
import {
  "Zy789kl/moon-change-point" @cp,
}

///|
fn main {
  let detector = @cp.Cusum::new(target_mean=0.0, control_limit=5.0, drift=0.5)
  let result = detector.update_result(2.0, index=1)
  println(result.summary())
}

DetectionResult exposes changed, score, confidence, direction, index, and evidence, so callers can choose their own storage and alerting policy.

CLI

The repository includes a small reproducible command-line program:

moon run cmd/main

It generates a fixed 512-sample mean-and-variance-shift scenario and prints a Markdown benchmark table for CUSUM, Robust-Z, and projection ensemble detection. The library API remains the primary integration surface for applications.

Architecture

The root package is organized by responsibility:

• types.mbt, stats.mbt, and window.mbt define public results, stable online statistics, and bounded windows.
• cusum.mbt, page_hinkley.mbt, bayesian.mbt, adaptive_detectors.mbt, and advanced_detectors.mbt implement online detectors.
• pipeline.mbt, stream_engine.mbt, time_windows.mbt, events.mbt, and alert_routing.mbt compose detectors into streaming workflows.
• offline_analysis.mbt, segment_quality.mbt, features.mbt, and rank_tests.mbt provide historical analysis and feature-level evidence.
• multivariate.mbt, matrix.mbt, and correlation.mbt cover vector and cross-series monitoring.
• quality.mbt, calibration.mbt, sampling.mbt, replay.mbt, explainability.mbt, and reporting.mbt support validation, reproducibility, and integration.
• production_*.mbt provides production-facing contracts, preprocessing, forecasting, monitoring, incident lifecycle, SLOs, rollout guardrails, canary control, and telemetry export.
• cmd/main is a runnable benchmark entry point; it does not add a second library abstraction.

Benchmark

The benchmark uses a deterministic 512-sample scenario with change point 256, baseline 10.0, mean shift 2.0, pre-change noise 0.15, post-change noise 0.7, and seed 20260818.

detector	detections	first detection	precision	recall	F1
CUSUM	77	258	0.012987012987012988	1	0.025641025641025647
Robust-Z	11	257	0.09090909090909091	1	0.16666666666666669
Projection ensemble	1	257	1	1	1

The complete table, checksums, reproduction command, and local wall-clock measurements are in BENCHMARK.md.

Tests

The repository contains boundary, regression, integration, and long-stream tests for the public API:

moon fmt --check
moon check --deny-warn --target all
moon build --target all
moon test --deny-warn --target all

The current test suite passes 425 tests. moon info regenerates the tracked public interface files so API changes are visible in review.

CI

GitHub Actions installs the latest stable MoonBit toolchain and runs format, warning-free checks, all-target builds, generated-interface checks, and tests on Ubuntu, macOS, and Windows. The reproducible benchmark workflow runs manually or for version tags.

License

Apache-2.0. See LICENSE.

The published package is available on Mooncakes as Zy789kl/moon-change-point@0.3.1, and the source repository is github.com/Zy789kl/moon-change-point.