Changelog

All notable changes to pickbuckets are documented here. The format follows Keep a Changelog and the project uses Semantic Versioning.

0.5.0 - 2026-06-03

Phase 5 — production and performance.

Experimental streaming bucketing under pickbuckets.experimental: StreamingEqualFrequencyBucket with chunked .partial_fit() and a bounded StreamingHistogram quantile sketch. Fitting is approximate; the produced rule applies through the same pure-Python runtime as the exact bucketers. Approximation error is bounded by max_centroids and converges to exact edges as it grows.
Optional matplotlib plotting helpers under pickbuckets.plotting (plot_bucket_counts, plot_target_rate, plot_woe), behind the pickbuckets[plot] extra. Each helper returns a matplotlib Axes for further customisation and never forces display.
Reproducible benchmark suite in benchmarks/run_benchmarks.py covering fit/transform across sizes, the runtime apply path versus adapter paths, and high-cardinality categorical grouping, with --json output for before/after comparisons. Each benchmark also reports peak heap memory via tracemalloc, including a bounded-memory check for large streaming inputs (--no-memory skips the memory pass).
API reference documentation in docs/api-reference.md, leading with the export → apply-without-the-stack portability demo.

The core package remains dependency-free; streaming is pure-Python and the no-heavy-deps CI gate is unchanged.

Phase 4 — advanced and supervised binning.

DecisionTreeBucket, WoEBucket (with WoE/IV, smoothing, minimum-bin-size, and monotonic constraints), ChiMergeBucket, and ExternalSplitBucket.
Supervised summaries (summary_table(), iv_summary()) that expose target rates and counts without leaking raw targets into serialized rules.

Phase 3 — real-world robustness.

Phase 2 — developer experience and integrations.

Public bucketer classes and scikit-learn-compatible adapters.
pandas Series/DataFrame support and first-class Polars Series/DataFrame/LazyFrame support.
AutoBucket smart type dispatch.

Phase 1 — core MVP and portable rule foundation.