qs

qs provides an interface for quickly saving and reading objects to and from disk. The goal of this package is to provide a lightning-fast and complete replacement for the saveRDS and readRDS functions in R.

Inspired by the fst package, qs uses a similar block-compression design using either the lz4 or zstd compression libraries. It differs in that it applies a more general approach for attributes and object references.

saveRDS and readRDS are the standard for serialization of R data, but these functions are not optimized for speed. On the other hand, fst is extremely fast, but only works on data.frame’s and certain column types.

qs is both extremely fast and general: it can serialize any R object like saveRDS and is just as fast and sometimes faster than fst.

Usage

library(qs)
df1 <- data.frame(x=rnorm(5e6), y=sample(5e6), z=sample(letters,5e6, replace=T))
qsave(df1, "myfile.qs")
df2 <- qread("myfile.qs")

Installation:

For R version 3.5 or higher:

# CRAN version
install.packages("qs")

# CRAN version compile from source with AVX2 support (recommended)
remotes::install_cran("qs", type="source", configure.args="--with-simd=AVX2")

# Experimental
remotes::install_github("traversc/qs", configure.args="--with-simd=AVX2")

For R version 3.4 and lower:

remotes::install_github("traversc/qs", ref = "qs34")

Features

The table below compares the features of different serialization approaches in R.

	qs	fst	saveRDS
Not Slow	✔	✔	❌
Numeric Vectors	✔	✔	✔
Integer Vectors	✔	✔	✔
Logical Vectors	✔	✔	✔
Character Vectors	✔	✔	✔
Character Encoding	✔	(vector-wide only)	✔
Complex Vectors	✔	❌	✔
Data.Frames	✔	✔	✔
On disk row access	❌	✔	❌
Attributes	✔	Some	✔
Lists / Nested Lists	✔	❌	✔
Multi-threaded	✔	✔	❌

qs also includes a number of advanced features:

For character vectors, qs also has the option of using the new alt-rep system (R version 3.5+) to quickly read in string data.
For numerical data (numeric, integer, logical and complex vectors) qs implements byte shuffling filters (adopted from the Blosc meta-compression library). These filters utilize extended CPU instruction sets (either SSE2 or AVX2).

Both of these features have the possibility of additionally increasing performance by orders of magnitude, for certain types of data. See sections below for more details.

Summary Benchmarks

The following benchmarks were performed on a Ryzen 2700x desktop using various data types (detailed below). qs was compared with saveRDS/readRDS in base R and the fst package for serializing and de-serializing a medium sized data.frame with 5 million rows (approximately 115 Mb):

data.frame(a=rnorm(5e6), 
           b=rpois(100,5e6),
           c=sample(starnames$IAU,5e6,T),
           d=sample(state.name,5e6,T),
           stringsAsFactors = F)

qs is highly parameterized and can be tuned by the user to extract as much speed and compression as possible, if desired. For simplicity, qs comes with 4 presets, which trades speed and compression ratio: “fast”, “balanced”, “high” and “archive”.

The tables and plots below summarize the performance of saveRDS, qs and fst with various parameters:

Summary table

Algorithm	Threads	Write Time (s)	Read Time (s)	File Size (Mb)
saveRDS / readRDS	1	4.680	1.500	55.2
saveRDS / readRDS	4	1.370	1.050	55.0
fst C=0	1	0.186	0.288	121.0
fst C=0	4	0.184	0.286	121.0
fst C=50	1	0.188	0.300	92.0
fst C=50	4	0.183	0.296	92.0
fst C=85	1	0.612	0.371	70.5
fst C=85	4	0.463	0.332	70.5
qs:lz4 shuffle=0 C=100 (fast)	1	0.196	0.319	106.0
qs:lz4 shuffle=0 C=100	4	0.161	0.322	106.0
qs:lz4 shuffle=7 C=1 (balanced)	1	0.262	0.363	59.4
qs:lz4 shuffle=7 C=1	4	0.194	0.365	59.4
qs:zstd shuffle=7 C=4 (high)	1	0.393	0.409	50.0
qs:zstd shuffle=7 C=4	4	0.212	0.411	50.0
qs:zstd_stream shuffle=7 C=14 (archive)	1	9.160	0.452	46.9

Serializing

df_bench_write

De-serializing

df_bench_read

Benchmarking write and read speed is a bit tricky and depends highly on a number of factors, such as operating system, the hardware being run on, the distribution of the data, or even the state of the R instance. Reading data is also further subjected to various hardware and software memory caches.

Generally speaking, qs and fst are considerably faster than saveRDS regardless of using single threaded or multi-threaded compression. qs also manages to achieve superior compression ratio through various optimizations (e.g. see “Byte Shuffle” section below).

Byte Shuffle

Byte shuffling (adopted from the Blosc meta-compression library) is a way of re-organizing data to be more ammenable to compression. For example: an integer contains four bytes and the limits of an integer in R are +/- 2^31-1. However, most real data doesn’t use anywhere near the range of possible integer values. For example, if the data were representing percentages, 0% to 100%, the first three bytes would be unused and zero.

Byte shuffling rearranges the data such that all of the first bytes are blocked together, the second bytes are blocked together, etc. This procedure often makes it very easy for compression algorithms to find repeated patterns and can often improves compression ratio by orders of magnitude. In the example below, shuffle compression achieves a compression ratio of over 1000x. See ?qsave for more details.

# With byte shuffling
x <- 1:1e8
qsave(x, "mydat.qs", preset="custom", shuffle_control=15, algorithm="zstd")
cat( "Compression Ratio: ", as.numeric(object.size(x)) / file.info("mydat.qs")$size, "\n" )
# Compression Ratio:  1389.164

# Without byte shuffling
x <- 1:1e8
qsave(x, "mydat.qs", preset="custom", shuffle_control=0, algorithm="zstd")
cat( "Compression Ratio: ", as.numeric(object.size(x)) / file.info("mydat.qs")$size, "\n" )
# Compression Ratio:  1.479294

Alt-rep character vectors

The alt-rep system was introduced in R version 3.5. Briefly, alt-rep vectors are objects that are not represented by R internal data, but have accesor functions which promise to “materialize” elements within the vector on the fly. To the user, this system is completely hidden and appears seamless.

In qs, only alt-rep character vectors are implemented because it is often the mostly costly of data types to read into R. Numeric and integer data are already fast enough and do not largely benefit. An example use case: if you have a large data.frame, and you are only interested in processing certain columns, it is wasted computation to materialize the whole data.frame. The alt-rep system solves this problem.

df1 <- data.frame(x = randomStrings(1e6), y = randomStrings(1e6), stringsAsFactors = F)
qsave(df1, "temp.qs")
rm(df1); gc() ## remove df1 and call gc for proper benchmarking

# With alt-rep
system.time(qread("temp.qs", use_alt_rep=T))[1]
#     0.109 seconds


# Without alt-rep
gc(verbose=F)
system.time(qread("temp.qs", use_alt_rep=F))[1]
#     1.703 seconds

qs