Fast serialization of R objects
Branch: master
Clone or download
Latest commit 73ce289 Feb 19, 2019

qs Build Status CRAN_Status_Badge

Quick serialization of R objects

This package provides an interface for quickly writing (serializing) and reading (de-serializing) 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 approach using the zstd library and direct "in memory" compression, which allows for lightning quick serialization. It differs in that it uses a more general approach for attributes and object references for common data types (numeric data, strings, lists, etc.), meaning any S3 object built on common data types, e.g., tibbles, time-stamps, bit64, etc. can be serialized. For less common data types (formulas, environments, functions, etc.), qs relies on built in R serialization functions via the RApiSerialize package followed by block-compression.

For character vectors, qs also uses the alt-rep system to quickly read in string data.


For R version 3.5 or higher:

install.packages("qs") or devtools::install_github("traversc/qs")

For R version 3.4 and lower:

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

Summary Benchmarks

The table below lists a more complete benchmark of serialization speeds for several different data types and methods.

qs saveRDS fst
1 thread
4 threads
Write Read Write Read Write Read Write Read
Integer Vector
900.5 MB/s
854.3 MB/s 27.1 MB/s 135.5 MB/s 1012.0 MB/s 421.8 MB/s 1111.6 MB/s 534.7 MB/s
Numeric Vector
938.4 MB/s 963.4 MB/s 24.3 MB/s 131.9 MB/s 992.2 MB/s 629.8 MB/s 1072.4 MB/s 795.3 MB/s
Character Vector
1446.8 MB/s 764.8 MB/s* 49.1 MB/s 43.9 MB/s 1663.5 MB/s 59.7 MB/s 1710.9 MB/s 58.8 MB/s
192.7 MB/s
273.1 MB/s 7.7 MB/s 123.5 MB/s N/A N/A N/A N/A
57.8 MB/s 120.7 MB/s 7.7 MB/s 89.6 MB/s N/A N/A N/A N/A


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

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

Additional Benchmarks

Data.Frame benchmark

Benchmarks for serializing and de-serializing large data.frames (5 million rows) composed of a numeric column (rnorm), an integer column (sample(5e6)), and a character vector column (random alphanumeric strings of length 50). See dataframe_bench.png for a comparison using different compression parameters.

This benchmark also includes materialization of alt-rep data, for an apples-to-apples comparison.

Serialization speed with default parameters:

Method write time (s) read time (s)
qs 0.49391294 8.8818166
fst (1 thread) 0.37411811 8.9309314
fst (4 thread) 0.3676273 8.8565951
saveRDS 14.377122 12.467517

Serialization speed with different parameters

The numbers in the figure reflect the compression parameter used. qs uses the zstd compression library, and compression parameters range from -50 to 22 (qs uses a default value of -1). fst defines it's own compression range through a combination of zstd and lz4 algorithms, ranging from 0 to 100 (default: 0).

Nested List benchmark

Benchmarks for serialization of random nested lists with random attributes (approximately 50 Mb). See the nested list example in the tests folder.

Serialization speed with default parameters:

Method write time (s) read time (s)
qs 0.17840716 0.19489372
saveRDS 3.484225 0.58762548