qs2 is the successor to the qs package. The goal is to have reliable and fast performance for saving and loading objects in R.

The qs2 format directly uses R serialization (via the R_Serialize/R_Unserialize C API) while improving underlying compression and disk IO patterns. If you are familiar with the qs package, the benefits and usage are the same.

Use the file extension qs2 to distinguish it from the original qs package. It is not compatible with the original qs format.

Installation

install.packages("qs2")

On Mac or Linux, you can enable multi-threading by compiling from source. It is enabled by default on Windows.

remotes::install_cran("qs2", type = "source", configure.args = " --with-TBB --with-simd=AVX2")

Multi-threading in qs2 uses the Intel Thread Building Blocks framework via the RcppParallel package.

Converting qs2 to RDS

Because the qs2 format directly uses R serialization, you can convert it to RDS and vice versa.

file_qs2 <- tempfile(fileext = ".qs2")
file_rds <- tempfile(fileext = ".RDS")
x <- runif(1e6)

# save `x` with qs_save
qs_save(x, file_qs2)

# convert the file to RDS
qs_to_rds(input_file = file_qs2, output_file = file_rds)

# read `x` back in with `readRDS`
xrds <- readRDS(file_rds)
stopifnot(identical(x, xrds))

Validating file integrity

The qs2 format saves an internal checksum. This can be used to test for file corruption before deserialization via the validate_checksum parameter, but has a minor performance penalty.

qs_save(data, "myfile.qs2")
data <- qs_read("myfile.qs2", validate_checksum = TRUE)

The qdata format

The package also introduces the qdata format which has its own serialization layout and works with only data types (vectors, lists, data frames, matrices).

It will replace internal types (functions, promises, external pointers, environments, objects) with NULL. The qdata format differs from the qs2 format in that it is NOT a general.

The eventual goal of qdata is to also have interoperability with other languages, particularly Python.

qd_save(data, "myfile.qs2")
data <- qd_read("myfile.qs2")

Benchmarks

A summary across 4 datasets is presented below.

Single-threaded

Algorithm	Compression	Save Time (s)	Read Time (s)
qs2	7.96	13.4	50.4
qdata	8.45	10.5	34.8
base::serialize	1.1	8.87	51.4
saveRDS	8.68	107	63.7
fst	2.59	5.09	46.3
parquet	8.29	20.3	38.4
qs (legacy)	7.97	9.13	48.1

Multi-threaded (8 threads)

Algorithm	Compression	Save Time (s)	Read Time (s)
qs2	7.96	3.79	48.1
qdata	8.45	1.98	33.1
fst	2.59	5.05	46.6
parquet	8.29	20.2	37.0
qs (legacy)	7.97	3.21	52.0

qs2, qdata and qs with compress_level = 3
parquet via the arrow package using zstd compression_level = 3
base::serialize with ascii = FALSE and xdr = FALSE

Datasets used

1000 genomes non-coding VCF 1000 genomes non-coding variants (2743 MB)
B-cell data B-cell mouse data, Greiff 2017 (1057 MB)
IP location IPV4 range data with location information (198 MB)
Netflix movie ratings Netflix ML prediction dataset (571 MB)

These datasets are openly licensed and represent a combination of numeric and text data across multiple domains. See inst/analysis/datasets.R on Github.