Skip to content

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

2.0.0

Added

lowess and fastLowess:

  • Added iterations_used: Option<usize> field to OnlineOutput<T>, reporting the number of robustness iterations performed when UpdateMode::Full is active. Returns Some(0) for the degenerate two-point linear fit and None when UpdateMode::Incremental is used.
  • Added ParseErrors(Vec<LowessError>) variant to LowessError, which collects all string-parse failures that accumulate in the builder and reports them together when build() is called.
  • Added "take_first" and "take_last" as accepted string aliases for MergeStrategy::TakeFirst and MergeStrategy::TakeLast.
  • Added "resmooth" as an accepted string alias for UpdateMode::Full and "single" as an alias for UpdateMode::Incremental, aligning string-parse behaviour with the loess-rs crate.
  • Added custom_weights(Vec<T>) builder method on LowessBuilder (Batch adapter only). Accepts a vector of non-negative per-observation weights that are multiplied into the distance and robustness weights before each local regression, allowing known-bad points to be suppressed (0.0) or high-quality measurements to be emphasised.
  • Centralized all impl FromStr blocks for the seven option enums (WeightFunction, BoundaryPolicy, ScalingMethod, RobustnessMethod, ZeroWeightFallback, MergeStrategy, UpdateMode) directly in api.rs, consolidating previously scattered implementations into a single source of truth. Parse and canonical-name helpers are exposed via lowess::internals::alias (requires dev feature), allowing fastLowess::binding_support to delegate all string-to-enum parsing through that path.
  • Added module-level defaults.rs files within each sub-module (math/, algorithms/, adapters/) to centralize default values close to the types they govern, propagating them from a single source of truth to ensure consistency across bindings and crates.

Python:

  • Added OnlineOutput class to the Python binding. OnlineLowess.add_point() now returns OnlineOutput | None instead of float | None, exposing smoothed, std_error, residual, robustness_weight, and iterations_used.
  • Added custom_weights parameter to the Lowess.fit(x, y, custom_weights=None) method. Accepts a list[float] of non-negative per-observation weights. Batch only.
  • Removed smooth(), smooth_streaming(), and smooth_online() convenience function stubs from _core.pyi.

R:

  • Added custom_weights parameter to Lowess$fit(). Accepts a numeric vector of non-negative per-observation weights. Batch only.

Julia:

  • Added custom_weights keyword argument to fit(model, x, y; custom_weights). Accepts a Vector{Float64} of non-negative per-observation weights. Batch only.

Node.js:

  • Added OnlineOutput object to the Node.js binding. OnlineLowess.add_point() now returns OnlineOutput | null instead of number | null, exposing smoothed, std_error, residual, robustness_weight, and iterations_used.
  • Added return_se and cv_seed fields to SmoothOptions.
  • Added customWeights as an optional per-call argument to fit(x, y, customWeights?) and fit_async(x, y, customWeights?). Accepts a Float64Array of non-negative per-observation weights. Includes pre-flight length-mismatch and non-negative validation. Batch only.
  • Added JavaScript-layer option key validation: unknown keys in SmoothOptions, StreamingOptions, or OnlineOptions now throw a TypeError listing all valid keys, via wrapper classes around the native NAPI exports.

WebAssembly:

  • Added custom_weights field to LowessOptions (passed in the options object to smooth()). Accepts a Float64Array of non-negative per-observation weights. Batch only.

C++:

  • Added custom_weights field to LowessOptions and a second overload of Lowess::fit() that accepts a const std::vector<double>& custom_weights argument. Values must be non-negative and length must match the input data. Batch only.

Changed

Monorepo:

  • Renamed all public API method and option names from camelCase to snake_case across every binding and all documentation. This is a breaking change for all consumers of the C++, Node.js, and WASM APIs.
  • Converted all documentation tables to compact single-space format.
  • Updated .clang-tidy to configure lower_case as the required naming convention for functions and member functions, matching the new snake_case public API.
  • Moved BENCHMARKS.md, CHANGELOG.md, and CONTRIBUTING.md from the repository root into docs/ and added them to the documentation site navigation.
  • Added a [patch.crates-io] section to the root Cargo.toml so all workspace bindings resolve fastLowess and lowess to the local workspace crates during development, replacing the previously-used registry (crates.io) versions.
  • Eliminated all local parse_* functions that each binding previously duplicated independently. Option parsing and builder application now delegates to fastLowess::binding_support, ensuring consistent aliases, validation messages, and behaviour across every language frontend.
  • Replaced direct use of KFold / LOOCV constructor types in the cross-validation path with binding_support::apply_cross_validation.
  • Split the Julia release CI workflow into two separate workflows: release-julia-jll.yml (triggered on release, opens the Yggdrasil PR) and release-julia-register.yml (manual dispatch, triggers JuliaRegistrator once the JLL PR is merged).
  • Major documentation improvements.

lowess and fastLowess:

  • Added Lowess<T>, StreamingLowess<T>, and OnlineLowess<T> type aliases as the primary user-facing constructors (e.g. StreamingLowess::new().chunk_size(50).build()). Mode-specific builder methods (chunk_size, overlap, window_capacity, min_points, update_mode) are now called directly on the type alias rather than after .adapter().
  • Made BatchLowessBuilder, StreamingLowessBuilder, and OnlineLowessBuilder internal-only: all public setter methods have been removed from these types. All smoothing configuration now flows through LowessBuilder<T, Mode> (exposed via the type aliases above). This is a breaking change for any code that called setter methods on an adapter builder directly.
  • Changed all enum-typed builder methods to accept strings instead: weight_function, robustness_method, scaling_method, boundary_policy, zero_weight_fallback, merge_strategy, and update_mode now take impl IntoEnum<T> (accepting both enum variants and strings such as .weight_function("tricube")) rather than requiring enum variants to be imported. This is a breaking change for any code passing enum variants directly.
  • Inlined the IntoEnum<E> trait and its macro-generated impls for all enum-typed builder parameters directly into api.rs (lowess) and binding_support.rs (fastLowess), eliminating a previously separate parse module. This allows builder methods to accept either a typed enum value (e.g. .weight_function(WeightFunction::Tricube)) or a string (e.g. .weight_function("tricube")) interchangeably.
  • Replaced the cross_validate(CVConfig) builder method (which required importing KFold or LOOCV types) with a string-based cross-validation API: .cv_method("kfold") / .cv_method("loocv"), .cv_k(n), .cv_fractions(vec![...]), and .cv_seed(n). KFold and LOOCV are no longer exported from the prelude. This is a breaking change for any code using the old cross_validate API.
  • Added a binding_support module providing shared helpers for all language binding frontends: string-to-enum parse functions (parse_weight_function, parse_robustness_method, parse_scaling_method, parse_boundary_policy, parse_zero_weight_fallback, parse_merge_strategy, parse_update_mode), matching canonical-string display functions, BuilderOptionSet / TypedBuilderOptionSet structs, and apply_builder_options / apply_typed_builder_options / apply_cross_validation helpers. This consolidates previously duplicated logic that was scattered across every binding into a single source of truth.
  • Renamed the internal auto_convergence struct field to auto_converge on BatchLowessBuilder, OnlineLowessBuilder, StreamingLowessBuilder, and the executor config types, making the field name consistent with the existing auto_converge() setter method. This is a breaking change for any code that accessed these fields directly.
  • Changed build() to wrap all accumulated string-parse errors in a LowessError::ParseErrors(Vec<LowessError>) value instead of surfacing only the first error. This is a breaking change for code that matched on LowessError::InvalidOption as the error returned from build().
  • Made the IntoEnum<E> trait pub(crate) in both lowess and fastLowess, restricting it to crate-internal use. Callers do not need to name this trait; builder methods continue to accept both enum variants and string literals unchanged.
  • Updated wide dependency to v1.5, wgpu to v30.0, and pollster to v1.0.

fastLowess:

  • Lowess, StreamingLowess, and OnlineLowess are now dedicated wrapper structs around LowessBuilder<f64> with string-accepting forwarding methods, rather than type aliases re-exported from the base lowess crate. Each wrapper's build() delegates to the corresponding parallel adapter and defaults to parallel execution. This is a breaking change: replace .adapter(Batch).build() with .build(), Lowess::new().adapter(Streaming) with StreamingLowess::new(), and Lowess::new().adapter(Online) with OnlineLowess::new().
  • The fastLowess prelude now exports only {Lowess, LowessError, LowessResult, OnlineLowess, StreamingLowess}, removing LowessBuilder, Adapter::{Batch, Online, Streaming}, and Backend::{CPU, GPU}. This is a breaking change for code that relied on those names being in scope via use fastLowess::prelude::*.

C++:

  • Renamed all public member functions to snake_case: make_error(), has_value(), r_squared(), effective_df(), residual_sd(), x_value(), y_value(), x_vector(), y_vector(), standard_errors(), confidence_lower(), confidence_upper(), prediction_lower(), prediction_upper(), robustness_weights(), fraction_used(), iterations_used(), process_chunk(), add_points().
  • Replaced Expected<LowessResult> OnlineLowess::add_points(const std::vector<double>&, const std::vector<double>&) with Expected<std::optional<double>> OnlineLowess::add_point(double x, double y). The method now processes a single point and returns only that point's smoothed value, or std::nullopt if not enough points have been accumulated yet. The underlying C FFI symbol is renamed from cpp_online_add_points to cpp_online_add_point. This is a breaking change.

Node.js:

  • Renamed all Diagnostics, SmoothOptions, StreamingOptions, and OnlineOptions interface fields to snake_case (r_squared, effective_df, residual_sd, chunk_size, merge_strategy, window_capacity, min_points, update_mode, and all smoothing option fields).
  • Renamed binding methods to snake_case: fit_async, process_chunk, add_points.
  • Renamed LowessResultObj getters to snake_case: standard_errors, confidence_lower, confidence_upper, prediction_lower, prediction_upper, robustness_weights, cv_scores, fraction_used, iterations_used.
  • Updated index.d.ts to reflect all renamed fields and methods.
  • Replaced add_points(x: Float64Array, y: Float64Array): LowessResultObj on OnlineLowess with add_point(x: number, y: number): OnlineOutput | null. The method now processes a single point and returns an OnlineOutput object, or null if not enough points have been accumulated yet. This is a breaking change.
  • Changed default OnlineOptions.window_capacity from 100 to 1000 and OnlineOptions.min_points from 2 to 3, matching the defaults used by the loess binding.
  • OnlineLowess now forwards all SmoothOptions fields to the underlying builder (previously only fraction, iterations, and parallel were forwarded; all other fields were hardcoded to None/false).
  • Updated napi-rs/cli dependency to v3.7 and oxlint to v1.73.

WASM:

  • Renamed all JS-facing option keys to snake_case by removing #[serde(rename = "camelCase")] attributes from SmoothOptions, StreamingOptions, and OnlineOptions. JSON passed from JavaScript must now use snake_case keys.
  • Updated Diagnostics getter names to snake_case: r_squared, effective_df, residual_sd.
  • Renamed the update(x: number, y: number) method on OnlineLowess to add_point(x: number, y: number). This is a breaking change.
  • Updated oxlint dependency to v1.73.

Python:

  • Renamed the update(x, y) method on OnlineLowess to add_point(x, y) and removed the separate array-based add_points(x, y) method. add_point processes a single point and returns the smoothed value as float | None. This is a breaking change.
  • Updated pyo3 and numpy dependencies to v0.29.

R:

  • Replaced $add_points(x, y) (vector inputs returning a list result) on OnlineLowess with $add_point(x, y) (scalar inputs returning numeric or NULL). The method now processes one point at a time and returns NULL until enough points have been accumulated. This is a breaking change.

Julia:

  • Replaced add_points(online, x::Vector{Float64}, y::Vector{Float64}) :: LowessResult with add_point(online, x::Float64, y::Float64) :: Union{Float64, Nothing}. The function now processes a single point and returns the smoothed value, or nothing if not enough points have been accumulated yet. The underlying C FFI symbol is renamed from jl_online_lowess_add_points to jl_online_lowess_add_point. This is a breaking change.

Fixed

C++:

  • Fixed remaining yVector() call in testBasicSmoothSerial that was missed during the snake_case rename (now y_vector()).

1.3.0

Added

Monorepo:

  • Added prerequisites for different bindings and platforms to CONTRIBUTING.md
  • Updated docs/assets/diagrams/lowess_smoothing_concept.svg to correctly illustrate LOWESS concepts (robustness iterations, bisquare re-weighting, outlier downweighting) instead of the generic LOESS algorithm it previously depicted.
  • Modified docs/requirements.txt to update the versions of the documentation dependencies.
  • Improved CI tests and coverage.
  • Modified Makefile to be truely cross-platform.
  • Added sanitizer check for all bindings and crates.

lowess:

  • Upgraded wide to version 1.4.

fastLowess:

  • Upgraded rayon to version 1.12.
  • Upgraded wgpu to version 29.0.

C++:

  • Added dedicated CMake packaging documentation in bindings/cpp/CMAKE.md for Windows installation, find_package(fastlowess CONFIG REQUIRED), and build-tree package discovery.

R:

  • Upgraded rextendr scaffold to 0.5.0: bumped Config/rextendr/version in DESCRIPTION and updated entrypoint.c to register the extendr panic hook (register_extendr_panic_hook()), so Rust panics now surface as R errors instead of crashing the session.
  • Added dev/fix_rd_style.R post-processing script to automatically normalize Rd file indentation (to 4 spaces) and wrap long lines (> 80 characters), ensuring compliance with CRAN/pkgcheck stylistic notes.
  • Added bindings/r/_pkgdown.yml configuration and updated the Makefile to use pkgdown::build_site(), satisfying the pkgcheck requirement for a dedicated documentation website.
  • Added automatic copying of shared tests from the project root into the R package within dev/prepare_cran.sh, ensuring R CMD build is self-contained.
  • Added direct extendr wrapper coverage tests plus extra validation-path tests in the R package, lifting covr::package_coverage() to 100% and clearing the package-level pkgcheck/goodpractice coverage complaints.

WASM:

  • Upgraded oxlint to 1.63.

Node.js:

  • Upgraded oxlint to 1.63.
  • Upgraded napi-rs/cli to 3.6.

Changed

lowess:

  • Updated MSRV to 1.89 to access the significant improvements made in wide since version 0.7.

fastLowess:

  • Updated MSRV to 1.89 to access the significant improvements made in wide since version 0.7.

R:

  • Refactored the R binding validation helpers to reuse validate_common_args() and coerce_nullable() in production code, split validate_params() into smaller helper validators, and consolidated duplicated constructor parameter documentation with @inheritParams before regenerating the Rd files.
  • Updated MSRV to 1.89 to access the significant improvements made in wide since version 0.7.
  • Abides by new rOpenSci standards.

Python:

  • Updated MSRV to 1.89 to access the significant improvements made in wide since version 0.7.

Node.js:

  • Updated MSRV to 1.89 to access the significant improvements made in wide since version 0.7.

Julia:

  • Updated MSRV to 1.89 to access the significant improvements made in wide since version 0.7.

WASM:

  • Updated MSRV to 1.89 to access the significant improvements made in wide since version 0.7.

C++:

  • Removed the legacy snake_case compatibility layer; the public C++ method API now uses camelBack, while variables and constants follow lower_case.
  • Updated MSRV to 1.89 to access the significant improvements made in wide since version 0.7.

Fixed

Monorepo:

  • Fixed R ASAN tests failing to compile vignettes by passing --no-build-vignettes to rcmdcheck.
  • Upgraded ASAN test environment to use modern rocker/r-devel-san:latest image and RDscript to resolve outdated readelf warnings.
  • Fixed Makefile idempotency checks on Linux by providing a default /tmp fallback for the TEMP directory variable.
  • Fixed accidental root Cargo.toml workspace isolation leaks by adding checked-in pre-commit and pre-push git hook guards that restore Cargo.toml.bak when present and fail loudly if required workspace members are still commented out.
  • Added a repo-local .cargo/config.toml that sets CC=clang-cl for x86_64-pc-windows-msvc, fixing Criterion 0.8 benchmark builds on Windows when cc-rs would otherwise pick clang.exe and fail to link alloca.

R:

  • Added strict pre-flight check for Pandoc in make r with a clear error message, as it is strictly required for building R Markdown vignettes on all platforms.
  • Fixed make r failure on Windows when the native tar.exe (bsdtar) doesn't support GNU-specific reproducible build flags by adding an automatic fallback.
  • Fixed FIND: Parameter format not correct error on Windows by replacing find commands with cross-platform rm -f wildcard expansions to prevent clashing with Windows' native find.exe.
  • Fixed local compilation error by automatically installing required Python TOML dependencies (tomli, tomli_w) in the Makefile's R build target.
  • Fixed make r vignette build crash by adding the missing BiocStyle R dependency to the Makefile's automated installation step.
  • Replaced the deprecated devtools::build_vignettes() command with pkgdown::build_articles() for local vignette previewing, and added pkgdown to the development dependencies.
  • Fixed cargo test and cargo build failing on Windows due to the default MSVC linker trying to link against R.lib by enforcing --target x86_64-pc-windows-gnu for all R-bound Cargo commands.
  • Fixed R CMD check WARNING on Windows caused by R's own Boolean.h using a C23 enum underlying type feature: the previous pragma-based suppression triggered a CRAN NOTE. Fixed by pre-defining a standard C Rboolean enum and setting the R_EXT_BOOLEAN_H_ header guard in entrypoint.c to prevent R's problematic version from loading.
  • Fixed R CMD check ERROR in the testing phase: test_check() always resolves tests from tests/testthat/ inside the package and ignores custom path arguments. Fixed by updating dev/prepare_cran.sh to copy shared tests from tests/r/testthat/ into bindings/r/tests/testthat/ before R CMD build, and simplifying testthat.R to a standard test_check("rfastlowess") call.
  • Fixed missing R packages (devtools, remotes) in the Makefile's dependency installation step.
  • Fixed %1 is not a valid Win32 application and section below image base DLL linkage errors on Windows by adding -Wl,--strip-all to Makevars.win PKG_LIBS.
  • Fixed rOpenSci pkgcheck warning by adding documentation website URL to DESCRIPTION.
  • Fixed CRAN note regarding non-API call R_NamespaceRegistry by upgrading extendr-api dependency to 0.9.0.
  • Fixed compilation error by providing the Result alias that was removed from extendr_api::prelude in 0.9.0.
  • Fixed the remaining SRR/pkgcheck findings in the R package by removing dead internal helper paths, reducing validate_params() cyclomatic complexity, eliminating duplicated roxygen parameter blocks, and covering the generated extendr-wrappers.R dispatch paths.

Python:

  • Fixed make python failing when ruff is not installed globally by bootstrapping ruff inside the Python virtual environment before formatting and linting.
  • Fixed make python on Windows by selecting the correct virtual environment activation script (.venv/Scripts/activate instead of the Unix-only .venv/bin/activate).
  • Fixed the Python public API to actually accept documented array-like inputs by coercing Lowess.fit(), StreamingLowess.process_chunk(), and OnlineLowess.add_points() arguments via np.asarray(..., dtype=np.float64) before calling the native extension.
  • Fixed Python wrapper analyzer issues by switching native extension lookups to runtime imports, avoiding wrapper class name shadowing in TYPE_CHECKING, and adding explicit wrapper docstrings.
  • Fixed false-positive Pylint warnings in bindings/python/python/fastlowess/_core.pyi by marking stub-only ellipsis bodies and signature arguments as intentional.

C++:

  • Fixed cbindgen idempotency check failure by adding automatic installation of the cbindgen CLI tool if missing.
  • Fixed explicit pointer checks, braces, named constants, and value-semantic result ownerships with compatibility wrappers.
  • Fixed all clang-tidy findings.
  • Fixed regenerated fastlowess.h clang-tidy regressions by normalizing the auto-generated header during the C++ bindings build, so FFI parameter naming and unused generated includes no longer come back after regeneration.
  • Fixed make cpp on Windows by making C++ symbol-export verification, CMake test execution, DLL runtime resolution, and Unix-specific test steps platform-aware.
  • Fixed MSVC size_t to unsigned long narrowing warnings in the C++ wrapper at the FFI boundary with explicit conversions.
  • Fixed C++ CMake package integration by generating and installing fastlowessConfig.cmake and related package export files for downstream find_package use.

fastLowess:

  • Fixed GPU execution under wgpu 29 by updating instance and pipeline layout setup, separating shader-written indirect dispatch data from the actual indirect dispatch buffer, stabilizing GPU buffer downloads, and correcting batched cross-validation dispatch offsets so the GPU integration test suite passes again.

Julia:

  • Fixed Windows local Julia runs by exporting an absolute FASTLOWESS_LIB path from the Makefile and moving DLL discovery in FastLOWESS.jl to runtime (__init__() plus runtime ccall), preventing stale precompiled library paths from being reused.
  • Linted the source code.

WASM:

  • Fixed deprecated JavaScript license-audit warnings by replacing the transient npx license-checker usage in the Makefile with a repo-local Node.js license summary script that still fails on GPL-family licenses.
  • Linted the source code.

Node.js:

  • Fixed make nodejs on Windows when /bin/bash could not launch npm from C:/Program Files/nodejs by using npm.cmd/npx.cmd in the Makefile.
  • Fixed deprecated JavaScript license-audit warnings by replacing the transient npx license-checker usage in the Makefile with a repo-local Node.js license summary script that still fails on GPL-family licenses.
  • Fixed .build() errors incorrectly using Status::GenericFailure; they now return Status::InvalidArg since build failures originate from invalid configuration (accumulated parse errors), not from runtime execution.
  • Linted the source code.

1.2.0

Added

R:

  • Added new tests.
  • Added a reference to the CONTRIBUTING.md file.
  • Added new examples for the print and plot methods.
  • Added test coverage evaluation.
  • Added missing srr tags.

Node.js:

  • Added advanced License Compliance check.
  • Added advanced dependency check.
  • Added advanced outdated dependency check.
  • Added advanced lock file check.
  • Added advanced TypeScript check.

WASM:

  • Added advanced License Compliance check.
  • Added advanced dependency check.
  • Added advanced outdated dependency check.
  • Added advanced lock file check.
  • Added WASM size check.

Changed

fastLowess:

  • Updated wgpu to v27.0 from v26.0.

Python:

  • Updated pyo3 to v0.28 from v0.27.
  • Updated numpy to v0.28 from v0.27.

R:

  • Removed the coerce_params dead code.
  • Spread srr tags to different files and removed extra tags from srr-stats-standards.R.

Node.js:

  • Switched from eslint to oxlint to remove vulnerabilities.

WASM:

  • Switched from eslint to oxlint to remove vulnerabilities.

Fixed

Monorepo:

  • Fixed project logo.

lowess:

  • Fixed documentation.
  • Fixed SRR tags.

Julia:

  • Linted examples.

Node.js:

  • Fixed vulnerabilities.

WASM:

  • Fixed vulnerabilities.
  • Fixed license.

C++:

  • Fixed memory leak in OnlineLowess.

1.1.2

Added

fastLowess:

  • Added srr tags

1.1.1

Added

lowess:

  • Added srr tags

Fixed

fastLowess:

  • Fixed memory layout mismatch in the GpuConfig struct
  • Refactored the GpuExecutor initialization in both the engine (gpu.rs) and tests (gpu_tests.rs) to handle missing hardware/drivers gracefully.
  • Improved the global executor lock handling to automatically recover from "poisoned" states. This prevents a single test crash from disabling the entire GPU backend for the remainder of the session.

1.1.0

Added

lowess:

  • Added Mean scaling method (Mean Absolute Deviation)
  • Added hooks for custom fitting backends
  • Added hooks for delegating boundary handling to the executor

fastLowess:

  • Added Mean scaling method (Mean Absolute Deviation)
  • Added support for different kernels to the GPU backend
  • Added support for different robustness methods to the GPU backend
  • Added support for different scaling methods to the GPU backend
  • Added support for different zero weight fallbacks to the GPU backend
  • Added support for different boundary policies to the GPU backend
  • Added support for auto convergence to the GPU backend
  • Added support for predictiona and confidence interval calculation to the GPU backend
  • Added support for cross-validation to the GPU backend

Fixed

lowess:

  • FitPassFn now returns Result to allow error propagation from custom fitting backends (e.g. GPU).
  • Adapters (Batch, Streaming, Online) now propagate errors from the executor instead of assuming success.
  • Fixed a bug where the Extend boundary policy was never applied.
  • Implemented Coordinate Centering to preserve precision during accumulation.

fastLowess:

  • Fixed potential integer overflow in GPU engine when dataset size exceeds u32::MAX.
  • Fixed panic in GPU initialization by propagating errors to the caller.
  • Fixed inefficient memory allocation in fit_all_points_tiled by reusing scratch buffers across tiles.
  • Fixed resource exhaustion in GPU backend by using a global Mutex for the executor instead of thread-local storage.

1.0.0

Added

Julia:

  • Added mean scaling method (Mean Absolute Deviation)

C++:

  • Added mean scaling method (Mean Absolute Deviation)

R:

  • Added mean scaling method (Mean Absolute Deviation)
  • Implemented print and plot methods for LowessResult objects
  • Added srr tags

Python:

  • Added mean scaling method (Mean Absolute Deviation)

Node.js:

  • Added mean scaling method (Mean Absolute Deviation)
  • Added JSDoc documentation to lib.rs for napi-rs generation
  • Added asynchronous support for batch processing

WASM:

  • Added mean scaling method (Mean Absolute Deviation)
  • Added an init_panic_hook function in src/lib.rs to be called by JS users during startup.
  • Added JSDoc documentation to lib.rs
  • Refactored the verbose Reflect::get boilerplate using serde and serde-wasm-bindgen. This allows us to define a Rust struct SmoothOptions and have wasm-bindgen automatically unpack the JS object into it.

Changed

lowess:

  • Refactored the constants to make the library robust safely against custom numeric types.
  • Minor improvements to the documentation.

C++:

  • Replaced exception-based error handling with a type-safe Expected<T> result type for all core methods (fit, process_chunk, finalize, add_points).
  • Refactored the internal FFI layer to use the idiomatic Rust From trait for converting result types.
  • Updated all C++ examples and tests to use the new Expected pattern, aligning the library with modern C++ practices.

Julia:

  • Wrapped all FFI functions in std::panic::catch_unwind. This ensures that if the Rust library panics (e.g., due to an internal assertion), it will be caught and reported as an error to Julia.

WASM:

  • Updated eslint/js, eslint, globals, and eslint-plugin-html packages to their latest versions.

Node.js:

  • Updated eslint/js, eslint, and globals packages to their latest versions.

Python:

  • Wrapped the heavy computation logic in py.allow_threads to allow Python to release the GIL during computation.

R:

  • Return results as LowessResult S3 objects instead of raw vectors

0.99.9

Changed

Monorepo:

  • Bump rust version to 1.88 for better stability
  • Change function-based builder pattern in the bindings to class-based builder pattern, allowing true streaming and online processing
  • Improve API docs

Julia:

  • Package is now registered on JuliaRegistries

C++:

  • Library is now available on conda-forge (libfastlowess)

R:

  • Package is now available on conda-forge (r-rfastlowess)

Node.js:

  • Package is now available on npm (fastlowess)

WASM:

  • Package is now available on npm (fastlowess-wasm)

0.99.8

Added

Julia:

  • Initial implementation

Node.js:

  • Initial implementation

WASM:

  • Initial implementation

C++:

  • Initial implementation

0.99.7

Changed

Python:

  • Switch to Stable ABI for CPython

Fixed

Monorepo:

  • Fix README file links
  • Fix Makefile bug with R versioning

0.99.6

Fixed

Monorepo:

  • Fix README file formats and links

0.99.5

Changed

Monorepo:

  • Reduced package size significantly by removing unnecessary dev files and docs from the final package.
  • Implemented comprehensive Cargo workspace inheritance pattern
  • Unified MSRV to 1.85.0
  • Centralized all metadata (version, authors, edition, license, etc.) in root Cargo.toml
  • All crates now use workspace = true for shared configuration
  • Created unified README.md for all crates/packages
  • Created unified CHANGELOG.md for all crates/packages
  • Created unified LICENSE for all crates/packages
  • Created unified .gitignore for all crates/packages
  • Added comprehensive badges from all packages

Fixed

lowess:

  • Fixed StreamingAdapter indexing bug that caused merged overlap points to be skipped in output
  • Simplified StreamingAdapter API: user now provides contiguous, non-overlapping chunks while the adapter handles internal buffering and merging
  • Standardized OnlineLowess default min_points to 2 (enabling smoothing after just one point)
  • Sanitized residual output to avoid "negative zero" (-0.0000) display for near-zero values

0.99.2

Changed

R:

  • Prepared package for rOpenSci Software Peer Review
  • Renamed main functions to avoid conflicts with base R:
  • smooth() → fastlowess()
  • smooth_online() → fastlowess_online()
  • smooth_streaming() → fastlowess_streaming()
  • Updated documentation to reflect new API and rOpenSci guidelines

Added

R:

  • Documentation website using pkgdown with automated deployment
  • Comprehensive function documentation with examples and cross-references
  • URL validation using urlchecker in CI workflow
  • Rigorous parameter validation for all exported functions
  • Expanded test suite achieving >96% coverage
  • Codecov CI workflow and badge

Fixed

R:

  • Documentation URLs
  • Package startup messages
  • pkgcheck workflow to run on host runner

0.99.1

Changed

R:

  • Modified package for Bioconductor submission

0.99.0

Added

R:

  • Support for new features in fastLowess v0.4.0, including NoBoundary boundary policy and MAD/MAR scaling methods

Changed

R:

  • Changed license from AGPL-3.0-or-later to dual MIT OR Apache-2.0
  • Updated documentation

0.7.0

Added

lowess:

  • NoBoundary variant to BoundaryPolicy enum (original Cleveland behavior)
  • ScalingMethod enum with MAR and MAD variants for configurable robust scale estimation
  • SIMD-optimized weighted least squares accumulation for f64 and f32
  • WLSSolver trait for type-specific SIMD dispatch
  • CVBuffer struct for pre-allocated cross-validation scratch buffers
  • VecExt trait for efficient vector reuse
  • Persistent scratch buffers to OnlineBuffer and StreamingBuffer

Changed

lowess:

  • Changed license from AGPL-3.0-or-later to dual MIT OR Apache-2.0
  • Refactored partition-related types
  • Replaced RangeInclusive iterations with while loops for improved performance
  • Optimized compute_window_weights and median_inplace
  • Added boundary thresholds for numerical stability
  • Unified scale estimation logic under ScalingMethod
  • Refactored LowessExecutor to accept optional external buffers
  • Optimized K-Fold Cross-Validation performance

0.6.0

Added

lowess:

  • cv_seed field to CVConfig for reproducible K-Fold cross-validation
  • Backend enum (CPU, GPU) as placeholder for GPU acceleration
  • Development-only fields: custom_fit_pass, custom_cv_pass, custom_interval_pass, backend, parallel
  • from_config and to_config methods to LowessExecutor

Changed

lowess:

  • Refactored cross_validate API to use CVConfig struct
  • Refactored Window::recenter to be bidirectional
  • Updated prelude to export enum variants directly
  • Reorganized src/engine/executor.rs into unified logical flow
  • Hidden internal-only fields from public documentation

Fixed

lowess:

  • Various broken documentation links
  • WeightParams struct to remove unused field
  • Bug in Batch and Streaming adapter conversion logic

Removed

lowess:

  • Unused GLSModel::local_wls method
  • CVMethod and CrossValidationStrategy enums
  • Type exports from prelude that caused ambiguity
  • .cargo/config.toml

0.5.3

Changed

lowess:

  • Consolidated validation logic into src/engine/validator.rs
  • Optimized sorting, window operations, MAD computation, and regression
  • Refactored robustness to use scratch buffers (allocation-free)
  • Optimized interpolation and cross-validation
  • Optimized delta interpolation with binary search

0.4.0

Added

fastLowess:

  • Zero-allocation parallel fitting via fit_all_points_parallel
  • Parallel CV memory reuse via cv_pass_parallel
  • Refined delta optimization for tied x-values
  • Parallel anchor precomputation for large datasets
  • Cache-oblivious tile-based processing

Python:

  • Support for new features in fastLowess v0.4.0

R:

  • Support for new features in fastLowess v0.4.0

Changed

lowess:

  • Transformed into core LOWESS implementation
  • Removed rayon and ndarray dependencies
  • Improved performance from 4-16× to 4-29× faster than statsmodels
  • Changed license from MIT to dual AGPL-3.0 and Commercial License
  • Reduced LOC from 3863 to 3263

fastLowess:

  • Changed license from AGPL-3.0-or-later to dual MIT OR Apache-2.0
  • Updated lowess dependency to v0.7.0
  • Implemented thread-local GpuExecutor persistence
  • Added intelligent buffer capacity management for GPU
  • Refactored GPU compute kernel with shared memory tiling

Python:

  • Changed license from AGPL-3.0-or-later to dual MIT OR Apache-2.0
  • Updated documentation

R:

  • Changed license from AGPL-3.0-or-later to dual MIT OR Apache-2.0
  • Updated documentation

Removed

lowess:

  • Validation and comparison code
  • Benchmarking code
  • Convenience re-exports

0.3.0

Added

fastLowess:

  • cpu (default) and gpu Cargo features
  • GPU execution engine in src/engine/gpu.rs
  • fit_pass_gpu function for GPU-accelerated processing
  • backend() setter method to all builders
  • Tests for GPU engine and parallel execution consistency

R:

  • Option to install from R-universe without Rust

Changed

lowess:

  • Updated Rust version to 1.86.0
  • Modified features: default std mode includes ndarray/std and rayon
  • Improved documentation

fastLowess:

  • Renamed builders: Extended*LowessBuilder → Parallel*LowessBuilder
  • Migrated parallel field to core lowess crate
  • Updated lowess dependency to v0.6.0
  • Made ndarray and rayon optional dependencies

Python:

  • Updated fastLowess dependency to v0.3.0
  • Refactored internal API usage
  • Updated cross-validation parameter handling

R:

  • Updated to fastLowess v0.3.0 and lowess v0.6.0
  • Updated cross-validation API

Fixed

lowess:

  • no-std build now compiles successfully

Python:

  • Documentation build errors
  • Bug where parallel argument was not exposed

R:

  • Automated vendor checksum fixing for CI builds

Removed

fastLowess:

  • .cargo/config.toml
  • Type exports from prelude that shadowed std types
  • Sequential, parallel, and ndarray adaptors

0.2.0

Added

Python:

  • Support for new features in fastLowess v0.2.0

R:

  • Support for new features in fastLowess v0.2.0

Changed

lowess:

  • Restructured project to reduce intra-module dependencies
  • Renamed "quartic" kernel to "biweight"
  • Cross-validation now uses true k-fold validation
  • Online LOWESS performs O(span) incremental updates
  • Numerous performance optimizations and numerical stability improvements

fastLowess:

  • Replaced linear scan with binary search in compute_anchor_points
  • Eliminated per-iteration division in interpolate_gap
  • Aligned with lowess crate v0.5.3 optimizations

Python:

  • Updated documentation
  • Changed module name from fastLowess to fastlowess

R:

  • Improved documentation

0.1.0

Added

lowess:

  • Initial LOWESS implementation based on Cleveland (1979)
  • Type-safe builder pattern API
  • Support for f32 and f64 types
  • Seven kernel weight functions
  • Statistical features (standard errors, confidence/prediction intervals)
  • Comprehensive diagnostics
  • Cross-validation with multiple strategies
  • Delta-based interpolation
  • Streaming and online processing variants
  • Optional parallel and ndarray features
  • Comprehensive error handling
  • Extensive documentation

fastLowess:

  • Initial release with parallel execution support

Python:

  • Python binding for fastLowess
  • Support for Python 3.14

R:

  • R binding for fastLowess