{disk.frame} has been soft-deprecated in favor of {arrow}. With the {arrow} 6.0.0 release, it’s now capable of doing larger-than-RAM data analysis quite well see release note. Hence, there is no strong reason to prefer {disk.frame} unless you have very specific feature needs.
For the above reason, I’ve decided to soft-deprecate {disk.frame} which means I will no longer actively develop new features for it but it will remain on CRAN in maintenance mode.
To help with the transition I’ve created a function, disk.frame::disk.frame_to_parquet(df, outdir)
to help you convert existing {disk.frame}s to the parquet format so you can use {arrow} with it.
I am working on an reincarnation of {disk.frame} in Julia, so the {disk.frame} will live on!
Thank your for support {disk.frame}. I’ve learnt alot along the way, but time has come to move on!
How do I manipulate tabular data that doesn’t fit into Random Access Memory (RAM)?
Use disk.frame!
In a nutshell, disk.frame makes use of two simple ideas
disk.frame performs a similar role to distributed systems such as Apache Spark, Python’s Dask, and Julia’s JuliaDB.jl for medium data which are datasets that are too large for RAM but not quite large enough to qualify as big data.
You can install the released version of disk.frame from CRAN with:
install.packages("disk.frame")
And the development version from GitHub with:
# install.packages("devtools")
devtools::install_github("xiaodaigh/disk.frame")
On some platforms, such as SageMaker, you may need to explicitly specify a repo like this
install.packages("disk.frame", repo="https://cran.rstudio.com")
Please see these vignettes and articles about disk.frame
{disk.frame}
which replicates the sparklyr
vignette for manipulating the nycflights13
flights data.{disk.frame}
which lists some commons way of creating disk.frames{disk.frame}
can be more epic! shows some ways of loading large CSVs and the importance of srckeep
dfglm
function for fitting generalized linear models{disk.frame}
and why create it?
disk.frame is an R package that provides a framework for manipulating larger-than-RAM structured tabular data on disk efficiently. The reason one would want to manipulate data on disk is that it allows arbitrarily large datasets to be processed by R. In other words, we go from “R can only deal with data that fits in RAM” to “R can deal with any data that fits on disk”. See the next section.
data.frame
and data.table
?
A data.frame
in R is an in-memory data structure, which means that R must load the data in its entirety into RAM. A corollary of this is that only data that can fit into RAM can be processed using data.frame
s. This places significant restrictions on what R can process with minimal hassle.
In contrast, disk.frame provides a framework to store and manipulate data on the hard drive. It does this by loading only a small part of the data, called a chunk, into RAM; process the chunk, write out the results and repeat with the next chunk. This chunking strategy is widely applied in other packages to enable processing large amounts of data in R, for example, see chunkded
arkdb
, and iotools
.
Furthermore, there is a row-limit of 2^31 for data.frame
s in R; hence an alternate approach is needed to apply R to these large datasets. The chunking mechanism in disk.frame provides such an avenue to enable data manipulation beyond the 2^31 row limit.
{disk.frame}
different to previous “big” data solutions for R?
R has many packages that can deal with larger-than-RAM datasets, including ff
and bigmemory
. However, ff
and bigmemory
restrict the user to primitive data types such as double, which means they do not support character (string) and factor types. In contrast, disk.frame makes use of data.table::data.table
and data.frame
directly, so all data types are supported. Also, disk.frame strives to provide an API that is as similar to data.frame
’s where possible. disk.frame supports many dplyr
verbs for manipulating disk.frame
s.
Additionally, disk.frame supports parallel data operations using infrastructures provided by the excellent future
package to take advantage of multi-core CPUs. Further, disk.frame uses state-of-the-art data storage techniques such as fast data compression, and random access to rows and columns provided by the fst
package to provide superior data manipulation speeds.
{disk.frame}
work?
disk.frame works by breaking large datasets into smaller individual chunks and storing the chunks in fst
files inside a folder. Each chunk is a fst
file containing a data.frame/data.table
. One can construct the original large dataset by loading all the chunks into RAM and row-bind all the chunks into one large data.frame
. Of course, in practice this isn’t always possible; hence why we store them as smaller individual chunks.
disk.frame makes it easy to manipulate the underlying chunks by implementing dplyr
functions/verbs and other convenient functions (e.g. the cmap(a.disk.frame, fn, lazy = F)
function which applies the function fn
to each chunk of a.disk.frame
in parallel). So that disk.frame can be manipulated in a similar fashion to in-memory data.frame
s.
{disk.frame}
different from Spark, Dask, and JuliaDB.jl?
Spark is primarily a distributed system that also works on a single machine. Dask is a Python package that is most similar to disk.frame, and JuliaDB.jl is a Julia package. All three can distribute work over a cluster of computers. However, disk.frame currently cannot distribute data processes over many computers, and is, therefore, single machine focused.
In R, one can access Spark via sparklyr
, but that requires a Spark cluster to be set up. On the other hand disk.frame requires zero-setup apart from running install.packages("disk.frame")
or devtools::install_github("xiaodaigh/disk.frame")
.
Finally, Spark can only apply functions that are implemented for Spark, whereas disk.frame can use any function in R including user-defined functions.
{disk.frame}
disk.frame works best if it can process multiple data chunks in parallel. The best way to set-up disk.frame so that each CPU core runs a background worker is by using
setup_disk.frame()
# this allows large datasets to be transferred between sessions
options(future.globals.maxSize = Inf)
The setup_disk.frame()
sets up background workers equal to the number of CPU cores; please note that, by default, hyper-threaded cores are counted as one not two.
Alternatively, one may specify the number of workers using setup_disk.frame(workers = n)
.
suppressPackageStartupMessages(library(disk.frame))
library(nycflights13)
# this will setup disk.frame's parallel backend with number of workers equal to the number of CPU cores (hyper-threaded cores are counted as one not two)
setup_disk.frame()
#> The number of workers available for disk.frame is 6
# this allows large datasets to be transferred between sessions
options(future.globals.maxSize = Inf)
# convert the flights data.frame to a disk.frame
# optionally, you may specify an outdir, otherwise, the
flights.df <- as.disk.frame(nycflights13::flights)
{disk.frame} aims to support as many dplyr verbs as possible. For example
flights.df %>%
filter(year == 2013) %>%
mutate(origin_dest = paste0(origin, dest)) %>%
head(2)
#> year month day dep_time sched_dep_time dep_delay arr_time
#> 1: 2013 1 1 517 515 2 830
#> 2: 2013 1 1 533 529 4 850
#> sched_arr_time arr_delay carrier flight tailnum origin dest
#> 1: 819 11 UA 1545 N14228 EWR IAH
#> 2: 830 20 UA 1714 N24211 LGA IAH
#> air_time distance hour minute time_hour origin_dest
#> 1: 227 1400 5 15 2013-01-01 05:00:00 EWRIAH
#> 2: 227 1416 5 29 2013-01-01 05:00:00 LGAIAH
Starting from disk.frame v0.3.0, there is group_by
support for a limited set of functions. For example:
result_from_disk.frame = iris %>%
as.disk.frame %>%
group_by(Species) %>%
summarize(
mean(Petal.Length),
sumx = sum(Petal.Length/Sepal.Width),
sd(Sepal.Width/ Petal.Length),
var(Sepal.Width/ Sepal.Width),
l = length(Sepal.Width/ Sepal.Width + 2),
max(Sepal.Width),
min(Sepal.Width),
median(Sepal.Width)
) %>%
collect
The results should be exactly the same as if applying the same group-by operations on a data.frame. If not, please report a bug.
If a function you like is missing, please make a feature request here. It is a limitation that function that depend on the order a column can only be obtained using estimated methods.
Function | Exact/Estimate | Notes |
---|---|---|
min |
Exact | |
max |
Exact | |
mean |
Exact | |
sum |
Exact | |
length |
Exact | |
n |
Exact | |
n_distinct |
Exact | |
sd |
Exact | |
var |
Exact |
var(x) only cor, cov support planned
|
any |
Exact | |
all |
Exact | |
median |
Estimate | |
quantile |
Estimate | One quantile only |
IQR |
Estimate |
library(data.table)
suppressWarnings(
grp_by_stage1 <-
flights.df[
keep = c("month", "distance"), # this analysis only required "month" and "dist" so only load those
month <= 6,
.(sum_dist = sum(distance)),
.(qtr = ifelse(month <= 3, "Q1", "Q2"))
]
)
#> data.table syntax for disk.frame may be moved to a separate package in the future
grp_by_stage1
#> qtr sum_dist
#> 1: Q1 27188805
#> 2: Q1 953578
#> 3: Q1 53201567
#> 4: Q2 3383527
#> 5: Q2 58476357
#> 6: Q2 27397926
The result grp_by_stage1
is a data.table
so we can finish off the two-stage aggregation using data.table syntax
grp_by_stage2 = grp_by_stage1[,.(sum_dist = sum(sum_dist)), qtr]
grp_by_stage2
#> qtr sum_dist
#> 1: Q1 81343950
#> 2: Q2 89257810
To find out where the disk.frame is stored on disk:
# where is the disk.frame stored
attr(flights.df, "path")
#> [1] "C:\\Users\\RTX2080\\AppData\\Local\\Temp\\Rtmpuc6nG8\\file1ad034f6ff6.df"
A number of data.frame functions are implemented for disk.frame
# get first few rows
head(flights.df, 1)
#> year month day dep_time sched_dep_time dep_delay arr_time
#> 1: 2013 1 1 517 515 2 830
#> sched_arr_time arr_delay carrier flight tailnum origin dest
#> 1: 819 11 UA 1545 N14228 EWR IAH
#> air_time distance hour minute time_hour
#> 1: 227 1400 5 15 2013-01-01 05:00:00
# get last few rows
tail(flights.df, 1)
#> year month day dep_time sched_dep_time dep_delay arr_time
#> 1: 2013 9 30 NA 840 NA NA
#> sched_arr_time arr_delay carrier flight tailnum origin dest
#> 1: 1020 NA MQ 3531 N839MQ LGA RDU
#> air_time distance hour minute time_hour
#> 1: NA 431 8 40 2013-09-30 08:00:00
# number of rows
nrow(flights.df)
#> [1] 336776
# number of columns
ncol(flights.df)
#> [1] 19
The work priorities at this stage are
Title | Language | Author | Date | Description |
---|---|---|---|---|
25 days of disk.frame | English | ZJ | 2019-12-01 | 25 tweets about disk.frame |
https://www.researchgate.net/post/What-is-the-Maximum-size-of-data-that-is-supported-by-R-datamining | English | Knut Jägersberg | 2019-11-11 | Great answer on using disk.frame |
{disk.frame} is epic |
English | Bruno Rodriguez | 2019-09-03 | It’s about loading a 30G file into disk.frame |
My top 10 R packages for data analytics | English | Jacky Poon | 2019-09-03 | disk.frame was number 3 |
useR! 2019 presentation video | English | Dai ZJ | 2019-08-03 | |
useR! 2019 presentation slides | English | Dai ZJ | 2019-08-03 | |
Split-apply-combine for Maximum Likelihood Estimation of a linear model | English | Bruno Rodriguez | 2019-10-06 | disk.frame used in helping to create a maximum likelihood estimation program for linear models |
Emma goes to useR! 2019 | English | Emma Vestesson | 2019-07-16 | The first mention of disk.frame in a blog post |
深入对比数据科学工具箱:Python3 和 R 之争(2020版) | Chinese | Harry Zhu | 2020-02-16 | Mentions disk.frame |
If you like disk.frame and want to speed up its development or perhaps you have a feature request? Please consider sponsoring disk.frame on Open Collective
{disk.frame}
Support disk.frame development by becoming a sponsor. Your logo will show up here with a link to your website.
Do you need help with machine learning and data science in R, Python, or Julia? I am available for Machine Learning/Data Science/R/Python/Julia consulting! Email me
Do you wish to give back the open-source community in non-financial ways? Here are some ways you can contribute
{disk.frame}
Github repo
{fst}
and {future}