Build Status CRAN_Status_Badge CRAN_Downloads_Badge CRAN_Downloads_Total_Badge

Quick serialization of R objects

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.


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


# CRAN version

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


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
On disk row access
Random column access
Attributes Some
Lists / Nested Lists

qs also includes a number of advanced features:

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 comparing qs, fst and saveRDS/readRDS in base R for serializing and de-serializing a medium sized data.frame with 5 million rows (approximately 115 Mb in memory):

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



(Benchmarks are based on qs ver. 0.21.2, fst ver. 0.9.0 and R 3.6.1.)

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. 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, and so on 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
system.time(qread("temp.qs", use_alt_rep=F))[1]
#     1.703 seconds

Serializing to memory

You can use qs to directly serialize objects to memory.


x <- qserialize(c(1,2,3))
[1] 1 2 3

Serializing objects to ASCII

The qs package includes two sets of utility functions for converting binary data to ASCII:

These functions are similar to base64 encoding functions found in various packages, but offer greater efficiency.


enc <- base91_encode(qserialize(datasets::mtcars, preset = "custom", compress_level = 22))
dec <- qdeserialize(base91_decode(enc))

(Note: base91 strings contain double quote characters (") and need to be single quoted if stored as a string.)

See the help files for additional details and history behind these algorithms.

Using qs within Rcpp

qs functions can be called directly within C++ code via Rcpp.

Example C++ script:

// [[Rcpp::depends(qs)]]
#include <Rcpp.h>
#include <qs.h>
using namespace Rcpp;

// [[Rcpp::export]]
void test() {
  qs::c_qsave(IntegerVector::create(1,2,3), "/tmp/myfile.qs", "high", "zstd", 1, 15, true, 1);

R side:

# save file using Rcpp interface
# read in file create through Rcpp interface
[1] 1 2 3

The C++ functions do not have default parameters; all parameters must be specified.

Future developments

Future versions will be backwards compatible with the current version.