A Suite of High-Performance Packages for Statistics and Data Manipulation.

fastverse

The fastverse is a suite of complementary high-performance packages for statistical computing and data manipulation in R. Developed independently by various people, fastverse packages jointly contribute to the objectives of:

• Speeding up R through heavy use of compiled code (C, C++, Fortran)
• Enabling more complex statistical and data manipulation operations in R
• Reducing the number of dependencies required for advanced computing in R

The fastverse package is a meta-package providing utilities for easy installation, loading and management of these packages. It is an extensible framework that allows users to (permanently) add or remove packages to create a 'verse' of packages suiting their general needs, or even create separate 'verses' of their own.

fastverse packages are jointly attached with library(fastverse), and several functions starting with fastverse_ help manage dependencies, detect namespace conflicts, add/remove packages from the fastverse and update packages. The vignette provides a concise overview of the package.

Core Packages

The fastverse installs with 4 core packages^[Before v0.3.0 matrixStats and fst were part of the core fastverse, but were removed following a poll in November 2022 which established that more than 50% of users don't use them actively.] (5 dependencies in total) which provide broad C/C++ based statistical and data manipulation functionality and have carefully managed APIs.

• data.table: Enhanced data frame class with concise data manipulation framework offering powerful aggregation, flexible split-apply-combine computing, reshaping, (rolling) joins, rolling statistics, set operations on tables, fast csv read/write, and various utilities such as transposition of data.

• collapse: Fast grouped and weighted statistical computations, time series and panel data transformations, list-processing, data manipulation functions, summary statistics and various utilities such as support for variable labels. Class-agnostic framework designed to work with vectors, matrices, data frames, lists and related classes including xts, data.table, tibble, plm, sf.

• kit: Parallel (row-wise) statistical functions, vectorized and nested switches, and some utilities such as efficient partial sorting.

• magrittr: Efficient pipe operators and aliases for enhanced R programming and code un-nesting.

Installation

# Install the CRAN version
install.packages("fastverse")

# Install (Windows/Mac binaries) from R-universe
install.packages("fastverse", repos = "https://fastverse.r-universe.dev")

# Install from GitHub (requires compilation)
remotes::install_github("fastverse/fastverse")

Note that the GitHub/r-universe version is not a development version, development takes place in the 'development' branch.

Extending the fastverse

Users can, via the fastverse_entend() function, freely attach extension packages. Setting permanent = TRUE adds these packages to the core fastverse. Another option is adding a .fastverse config file with packages to the project directory. Separate verses can be created with fastverse_child(). See the vignette for details.

Suggested Extensions

High-performing packages for different data manipulation and statistical computing topics are suggested below. The total (recursive) dependency count is indicated for each package.

Time Series

• xts and zoo: Fast and reliable matrix-based time series classes providing fully identified ordered observations and various utilities for plotting and computations (1 dependency).

• roll: Fast rolling and expanding window functions for vectors and matrices (3 dependencies).

  Notes: xts/zoo objects are preserved by roll functions and by collapse's time series and data transformation functions^[collapse functions can also handle irregular time series.]. As xts/zoo objects are matrices, all matrixStats functions apply to them as well. xts objects can also easily be converted to and from data.table, which also has some fast rolling functions like frollmean and frollapply.

Dates and Times

• anytime: Anything to 'POSIXct' or 'Date' converter (2 dependencies).

• fasttime: Fast parsing of strings to 'POSIXct' (0 dependencies).

• nanotime: Provides a coherent set of temporal types and functions with nanosecond precision -
  based on the 'integer64' class (7 dependencies).

• clock: Comprehensive library for date-time manipulations using a new family of orthogonal date-time classes (durations, time points, zoned-times, and calendars) (6 dependencies).

• timechange: Efficient manipulation of date-times accounting for time zones and daylight saving times (1 dependency).

  Notes: Date and time variables are preserved in many data.table and collapse operations. data.table additionally offers an efficient integer based date class 'IDate' with some supporting functionality. xts and zoo also provide various functions to transform dates, and zoo provides classes 'yearmon' and 'yearqtr' for convenient computation with monthly and quarterly data. Package mondate also provides a class 'mondate' for monthly data. Many users also find lubridate convenient for 'POSIX-' and 'Date' based computations.

Strings

• stringi: Main R package for fast, correct, consistent, and convenient string/text manipulation (backend to stringr and snakecase) (0 dependencies).

• stringfish: Fast computation of common (base R) string operations using the ALTREP system (2 dependencies).

• stringdist: Fast computation of string distance metrics, matrices, and fuzzy matching (0 dependencies).

  Notes: At least two packages offer convenient wrappers around the rather rich stringi API: stringr provides simple, consistent wrappers for common string operations, based on stringi (3 dependencies), and snakecase converts strings into any case, based on stringi and stringr (4 dependencies).

Statistics and Computing

• matrixStats: Efficient row-and column-wise (weighted) statistics on matrices and vectors, including computations on subsets of rows and columns (0 dependencies).

• Rfast and Rfast2: Heterogeneous sets of fast functions for statistics, estimation and data manipulation operating on vectors and matrices (4-5 dependencies).

• vctrs: Computational backend of the tidyverse that provides many basic programming functions for R vectors (including lists and data frames) implemented in C (such as sorting, matching, replicating, unique values, concatenating, splitting etc. of vectors). These are often significantly faster than base R equivalents, but generally not as aggressively optimized as some equivalents found in collapse or data.table (4 dependencies).

• parallelDist: Multi-threaded distance matrix computation (3 dependencies).

• coop: Fast implementations of the covariance, correlation, and cosine similarity (0 dependencies).

• rsparse: Implements many algorithms for statistical learning on sparse matrices - matrix factorizations, matrix completion, elastic net regressions, factorization machines (8 dependencies). See also package MatrixExtra.

• fastmatrix provides a small set of functions written in C or Fortran providing fast computation of some matrices and operations useful in statistics (0 dependencies).

• matrixTests efficient execution of multiple statistical hypothesis tests on rows and columns of matrices (1 dependency).

• rrapply: The rrapply() function extends base rapply() by including a condition or predicate function for the application of functions and diverse options to prune or aggregate the result (0 dependencies).

• dqrng: Fast uniform, normal or exponential random numbers and random sampling (i.e. faster runif, rnorm, rexp, sample and sample.int functions) (3 dependencies).

• fastmap: Fast implementation of data structures based on C++, including a key-value store (fastmap), stack (faststack), and queue (fastqueque) (0 dependencies).

• fastmatch: A faster match() function (drop-in replacement for base::match, and base::%in%), that keeps the hash table in memory for much faster repeated lookups (0 dependencies).

• hutilscpp provides C++ implementations of some frequently used utility functions in R (4 dependencies).

  Notes: Rfast has a number of like-named functions to matrixStats. These are simpler but typically faster and support multi-threading. Some highly efficient statistical functions can also be found scattered across various other packages, notable to mention here are Hmisc (60 dependencies) and DescTools (17 dependencies).

Spatial

• sf: Leading framework for geospatial computing and manipulation in R, offering a simple and flexible spatial data frame and supporting functionality (12 dependencies).

• s2: Provides R bindings for Google's s2 C++ library for high-performance geometric calculations on the sphere (3D, geographic/geodetic CRS). Used as a backend to sf for calculations on geometries with geographic/geodetic CRS, but using s2 directly can provide substantial performance gains (2 dependencies).

• geos: Provides an R API to the Open Source Geometry Engine (GEOS) C-library, which can be used to very efficiently manipulate planar (2D/flat/projected CRS) geometries, and a vector format with which to efficiently store 'GEOS' geometries. Used as a backend to sf for calculations on geometries with projected CRS, but using geos directly can provide substantial performance gains (2 dependencies).

• stars: Spatiotemporal data (raster and vector) in the form of dense arrays, with space and time being array dimensions (16 dependencies).

• terra: Methods for spatial data analysis with raster and vector data. Processing of very large (out of memory) files is supported (1 dependency).

• exactextractr: Provides fast extraction from raster datasets using polygons. Notably, it is much faster than terra for computing summary statistics of raster layers within polygons (17 dependencies).

• geodist: Provides very fast calculation of geodesic distances (0 dependencies).

• dggridR: Provides discrete global grids for R: allowing accurate partitioning of the earths surface into equally sized grid cells of different shapes and sizes (30 dependencies). The source project is not well maintained, and users are strongly encouraged to install this fork (version 3.1+) which fixes a major bug on Mac and introduces a collapse backend for faster grid materialization (14 dependencies).

• cppRouting: Algorithms for routing and solving the traffic assignment problem, including calculation of distances, shortest paths and isochrones on weighted graphs using several (optimized) variants of Dijkstra's algorithm (4 dependencies).

• igraph: Provides and R port to the igraph C library for complex network analysis and graph theory (11 dependencies).

  Notes: collapse can be used for efficient manipulation and computations on sf data frames. sf also offers tight integration with dplyr. Another efficient routing package is dodgr (45 dependencies). sfnetworks allows network analysis combining sf and igraph (42 dependencies) and functions for network cleaning (partly taken from tidygraph which also wraps igraph). stplanr facilitates sustainable transport planning with R, including very useful helpers such as overline() to turn a set of linestrings (routes) into a network (45 dependencies).

Visualization

• dygraphs: Interface to 'Dygraphs' interactive time series charting library (12 dependencies).

• lattice: Trellis graphics for R (0 dependencies).

• grid: The grid graphics package (0 dependencies).

• tinyplot provides a lightweight extension of the base R graphics system, with support for automatic grouping, legends, facets, and various other enhancements (0 dependencies).

• ggplot2: Create elegant data visualizations using the Grammar of Graphics (27 dependencies).

• scales: Scale functions for visualizations (11 dependencies).

  Notes:latticeExtra provides extra graphical utilities base on lattice. gridExtra provides miscellaneous functions for grid graphics (and consequently for ggplot2 which is based on grid). gridtext provides improved text rendering support for grid graphics. Many packages offer ggplot2 extensions, (typically starting with 'gg') such as ggExtra, ggalt, ggforce, ggh4x, ggmap, ggtext, ggthemes, ggrepel, ggridges, ggfortify, ggstatsplot, ggeffects, ggsignif, GGally, ggcorrplot, ggdendro, etc.. Users in desperate need for greater performance may also find the (unmaintained) lwplot package useful that provides a faster and lighter version of ggplot2 with data.table backend.

Data Manipulation in R Based on Faster Languages

• r-polars provides an R-port to the impressively fast polars DataFrame's library written in Rust (1 dependencies).

  Notes: Package tidypolars provides a tidyverse-style wrapper around r-polars.

Data Input-Output, Serialization, and Larger-Than-Memory Processing (IO)

• fst: A compressed data file format that is very fast to read and write. Full random access in both rows and columns allows reading subsets from a '.fst' file (2 dependencies).

• qs provides a lightning-fast and complete replacement for the saveRDS and readRDS functions in R. It supports general R objects with attributes and references - at similar speeds to fst - but does not provide on-disk random access to data subsets like fst (4 dependencies).

• arrow provides both a low-level interface to the Apache Arrow C++ library (a multi-language toolbox for accelerated data interchange and in-memory processing) including fast reading / writing delimited files, efficient storage of data as .parquet or .feather files, efficient (lazy) queries and computations, and sharing data between R and Python (14 dependencies). It provides methods for several dplyr functions allowing highly efficient data manipulation on arrow datasets. Check out the useR2022 workshop on working with larger than memory data with apache arrow in R, and the apache arrow R cookbook as well as the awesome-arrow-r repository.

• duckdb: DuckDB is a high-performance analytical database system that can be used on in-memory or out-of memory data (including csv, .parquet files, arrow datasets, and it's own .duckdb format), and that provides a rich SQL dialect and optimized query execution for data analysis (1 dependency). It can also be used with the dbplyr package that translates dplyr code to SQL. This Article by Christophe Nicault (October 2022) demonstrates the integration of duckdb with R and arrow. Also see the official docs.

• vroom provides fast reading of delimited files (23 dependencies).

  Notes: data.table provides fread and fwrite for fast reading of delimited files.

Compiling R

• nCompiler: Compiles R functions to C++, and covers basic math, distributions, vectorized math and linear algebra, as well as basic control flow. R and Compiled C++ functions can also be jointly utilized in the a class 'nClass' that inherits from R6. An in-progress user-manual provides an overview of the package.

• ast2ast: Also compiles R functions to C++, and is very straightforward to use (it has a single function translate() to compile R functions), but less flexible than nCompiler (e.g. it currently does not support linear algebra). Available on CRAN (6 dependencies).

• odin: Implements R to C translation and compilation, but specialized for differential equation solving problems. Available on CRAN (8 dependencies).

• armacmp translates linear algebra code written in R to C++ using the Armadillo Template Library. The package can also be used to write mathematical optimization routines that are translated and optimized in C++ using RcppEnsmallen.

• r2c provides compilation of R functions to be applied over many groups (e.g. grouped bivariate linear regression etc.).

• FastR is a high-performance implementation of the entire R programming language, that can JIT compile R code to run on the Graal VM.

• inline allows users to write C, C++ or Fortran functions and compile them directly to an R function for use within the R session. Available on CRAN (0 dependencies).

  Notes: Many of these projects are experimental and not available as CRAN packages.

R-like Data Manipulation in Faster Languages

• tidypolars is a python library built on top of polars that gives access to methods and functions familiar to R tidyverse users.

• Tidier.jl provides a Julia implementation of the tidyverse mini-language in Julia. Powered by the DataFrames.jl library.

R Bindings to Faster Languages

• R's C API is the most natural way to extend R and does not require additional packages. It is further documented in the Writing R Extensions Manual, the R Internals Manual, the r-internals repository and sometimes referred to in the R Blog (and some other Blogs on the web). Users willing to extend R in this way should familiarize themselves with R's garbage collection and PROTECT Errors.

• Rcpp provides seamless R and C++ integration, and is widely used to extend R with C++. Compared to the C API compile time is slower and object files are larger, but users don't need to worry about garbage collection and can use modern C++ as well as a rich set of R-flavored functions and classes (0 dependencies).

• cpp11 provides a simpler, header-only R binding to C++ that allows faster compile times and several other enhancements (0 dependencies).

• tidyCpp provides a tidy C++ wrapping of the C API of R - to make the C API more amenable to C++ programmers (0 dependencies).

• JuliaCall Provides an R interface to the Julia programming language (11 dependencies). Other interfaces are provided by XRJulia (2 dependencies) and JuliaConnectoR (0 dependencies).

• rextendr provides an R interface to the Rust programming language (29 dependencies).

• rJava provides an R interface to Java (0 dependencies).

  Notes: There are many Rcpp extension packages binding R to powerful C++ libraries, such as linear algebra through RcppArmadillo and RcppEigen, thread-safe parallelism through RcppParallel etc.

Tidyverse-like Data Manipulation built on data.table

• tidytable: A tidy interface to data.table that is rlang compatible. Quite comprehensive implementation of dplyr, tidyr and purr functions. Package uses a class tidytable that inherits from data.table. The dt() function makes data.table syntax pipeable (12 total dependencies).

• dtplyr: A tidy interface to data.table built around lazy evaluation i.e. users need to call as.data.table(), as.data.frame() or as_tibble() to access the results. Lazy evaluation holds the potential of generating more performant data.table code (20 dependencies).

• tidyfst: Tidy verbs for fast data manipulation. Covers dplyr and some tidyr functionality. Functions have _dt suffix and preserve data.table object. A cheatsheet is provided (7 dependencies).

• tidyft: Tidy verbs for fast data operations by reference. Best for big data manipulation on out of memory data using facilities provided by fst (7 dependencies).

• tidyfast: Fast tidying of data. Covers tidyr functionality, dt_ prefix, preserves data.table object (2 dependencies).

• maditr: Fast data aggregation, modification, and filtering with pipes and data.table. Minimal implementation with functions let() and take() for most common data manipulation tasks. Also provides Excel-like lookup functions (2 dependencies).

• table.express also o builds data.table expressions from dplyr verbs, without executing them eagerly. Similar to dtplyr but less mature (17 dependencies).

  Notes: These packages are wrappers around data.table and do not introduce own compiled code.

Parallelization, High-Performance Computing and Out-Of-Memory Data

• See the High-Performance and Parallel Computing Task View and the futureverse.

Adding to this list

Please notify me of any other packages you think should be included here. Such packages should be well designed, top-performing, low-dependency, and, with few exceptions, provide own compiled code. Please note that the fastverse focuses on general purpose statistical computing and data manipulation, thus I won't include fast packages to estimate specific kinds of models here (of which R also has a great many).