Crate driftwatch 

driftwatch — data & model drift detection for Rust.

driftwatch computes how far a live data or prediction distribution has moved from a trusted reference (training/baseline) distribution, and lets you alert on it. It provides the metrics named in the Rust ML ecosystem gap that Evidently/WhyLabs fill in Python — but no equivalent existed in Rust:

• Binned distribution metrics — psi, kl_divergence, js_divergence over Histograms.
• Raw-sample distributional tests — a two-sample ks_test and a categorical chi_square_test, each returning a statistic and a p-value so you apply your own significance threshold.
• Orchestration — ReferenceDistribution per feature, a DatasetMonitor that turns a live batch into a DriftReport, and a thread-safe LiveWindow for feeding it from a serving system.
• Pluggable alerting — the Alerter trait, with optional tracing and webhook implementations behind feature flags.

What this crate is not

It does not render HTML reports or a dashboard UI (pair DriftReport’s structured data with plotters-statistical for charts), it does not do general data-quality profiling beyond drift, and it recomputes drift over discrete windows/snapshots rather than as a continuously-updated online statistic. See the README for the full comparison against Evidently/WhyLabs.

Quick start

use driftwatch::{DatasetMonitor, ReferenceDistribution, EqualFrequencyBinning, LiveFeature};

// Fit a reference distribution per feature from baseline data.
let baseline: Vec<f64> = (0..200).map(|i| i as f64 / 200.0).collect();
let reference = ReferenceDistribution::fit_continuous(
    "score",
    &baseline,
    EqualFrequencyBinning::new(10).unwrap(),
)
.unwrap();

let mut monitor = DatasetMonitor::new();
monitor.add_feature(reference);

// Check a live batch that has shifted upward.
let live: Vec<f64> = (0..200).map(|i| 0.5 + i as f64 / 200.0).collect();
let report = monitor.check(&[("score", LiveFeature::Continuous(&live))]).unwrap();
println!("{report}");
assert!(report.dataset_drift_detected());

Re-exports

pub use binning::BinDefinition;

pub use binning::ContinuousBinning;

pub use binning::EqualFrequencyBinning;

pub use binning::EqualWidthBinning;

pub use binning::Histogram;

pub use binning::DEFAULT_BIN_COUNT;

pub use distribution::FeatureKind;

pub use distribution::LiveFeature;

pub use distribution::ReferenceDistribution;

pub use error::DriftError;

pub use error::Result;

pub use metrics::chi_square_test;

pub use metrics::js_divergence;

pub use metrics::kl_divergence;

pub use metrics::ks_test;

pub use metrics::psi;

pub use metrics::ChiSquareResult;

pub use metrics::KsTestResult;

pub use metrics::DEFAULT_EPSILON;

pub use monitor::DatasetMonitor;

pub use monitor::DriftReport;

pub use monitor::DriftVerdict;

pub use monitor::FeatureConfig;

pub use monitor::FeatureDrift;

pub use monitor::LiveWindow;

pub use monitor::MetricKind;

pub use monitor::MetricScore;

pub use monitor::PredictionDriftMonitor;

pub use monitor::WindowMode;

pub use monitor::LabelDriftMonitor;

pub use monitor::LabelDriftReport;

pub use dashboard::Dashboard;

pub use streaming::OnlineDistribution;

pub use streaming::PageHinkleyChange;

pub use streaming::PageHinkleyDetector;

pub use streaming::StreamingMonitor;

pub use streaming::StreamingReport;

pub use profile::CategoricalProfile;

pub use profile::ContinuousProfile;

pub use profile::DatasetProfile;

pub use profile::FeatureProfile;

pub use profile::Schema;

pub use profile::ValidationIssue;

pub use profile::ValidationReport;

pub use report::HtmlReport;

pub use alert::Alerter;

pub use alert::DriftAlertEvent;

pub use alert::NopAlerter;

pub use alert::LogAlerter;

pub use alert::WebhookAlerter;

Modules

alert

Pluggable alerting on drift detection.

binning

Binning: turning raw feature samples into comparable discretized distributions.

dashboard

A live, auto-refreshing drift dashboard (feature dashboard).

distribution

Per-feature reference distributions and the live data compared against them.

error

Error types for the crate.

export

Prometheus / metrics export (feature prometheus-export).

metrics

Drift metrics.

monitor

Orchestration: turning per-feature reference distributions into a dataset-level DriftReport, plus the LiveWindow buffer that feeds a monitor from a running service.

profile

Data-quality profiling and schema validation.

report

Static, self-contained HTML drift reports.

streaming

Continuously-updated online drift (feature streaming).