In tidyverse/multidplyr: A Multi-Process 'dplyr' Backend

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/README-",
  out.width = "100%"
)

multidplyr

Overview

multidplyr is a backend for dplyr that partitions a data frame across multiple cores. You tell multidplyr how to split the data up with partition() and then the data stays on each node until you explicitly retrieve it with collect(). This minimises the amount of time spent moving data around, and maximises parallel performance. This idea is inspired by partools by Norm Matloff and distributedR by the Vertica Analytics team.

Due to the overhead associated with communicating between the nodes, you won't see much performance improvement with simple operations on less than ~10 million observations, and you may want to instead try dtplyr, which uses data.table. multidplyr's strength is found parallelising calls to slower and more complex functions.

(Note that unlike other packages in the tidyverse, multidplyr requires R 3.5 or greater. We hope to relax this requirement in the future.)

Installation

You can install the released version of multidplyr from CRAN with:

install.packages("multidplyr")

And the development version from GitHub with:

# install.packages("pak")
pak::pak("tidyverse/multidplyr")

Usage

To use multidplyr, you first create a cluster of the desired number of workers. Each one of these workers is a separate R process, and the operating system will spread their execution across multiple cores:

library(multidplyr)

cluster <- new_cluster(4)
cluster_library(cluster, "dplyr")

There are two primary ways to use multidplyr. The first, and most efficient, way is to read different files on each worker:

# Create a filename vector containing different values on each worker
cluster_assign_each(cluster, filename = c("a.csv", "b.csv", "c.csv", "d.csv"))

# Use vroom to quickly load the csvs
cluster_send(cluster, my_data <- vroom::vroom(filename))

# Create a party_df using the my_data variable on each worker
my_data <- party_df(cluster, "my_data")

Alternatively, if you already have the data loaded in the main session, you can use partition() to automatically spread it across the workers. Before calling partition(), it's a good idea to call group_by() to ensure that all of the observations belonging to a group end up on the same worker.

library(nycflights13)

flight_dest <- flights %>% group_by(dest) %>% partition(cluster)
flight_dest

Now you can work with it like a regular data frame, but the computations will be spread across multiple cores. Once you've finished computation, use collect() to bring the data back to the host session:

flight_dest %>% 
  summarise(delay = mean(dep_delay, na.rm = TRUE), n = n()) %>% 
  collect()

Note that there is some overhead associated with copying data from the worker nodes back to the host node (and vice versa), so you're best off using multidplyr with more complex operations. See vignette("multidplyr") for more details.

tidyverse/multidplyr documentation built on Nov. 7, 2023, 12:55 a.m.