qs2: a framework for efficient serialization
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.
qs_save(data, "myfile.qs2")
data <- qs_read("myfile.qs2")
Use the file extension qs2
to distinguish it from the original qs
package. It is not compatible with the original qs
format.
install.packages("qs2")
On x64 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")
On non-x64 systems (e.g. Mac ARM) remove the AVX2 flag.
remotes::install_cran("qs2", type = "source", configure.args = "--with-TBB")
Multi-threading in qs2
uses the Intel Thread Building Blocks
framework via the RcppParallel
package.
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))
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 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")
A summary across 4 datasets is presented below.
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 |
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
andqs
withcompress_level = 3
parquet
via thearrow
package using zstdcompression_level = 3
base::serialize
withascii = FALSE
andxdr = 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.
Serialization functions can be accessed in compiled code. Below is an example using Rcpp.
// [[Rcpp::depends(qs2)]]
#include <Rcpp.h>
#include "qs2_external.h"
using namespace Rcpp;
// [[Rcpp::export]]
SEXP test_qs_serialize(SEXP x) {
size_t len = 0;
unsigned char * buffer = c_qs_serialize(x, &len, 10, true, 4); // object, buffer length, compress_level, shuffle, nthreads
SEXP y = c_qs_deserialize(buffer, len, false, 4); // buffer, buffer length, validate_checksum, nthreads
c_qs_free(buffer); // must manually free buffer
return y;
}
// [[Rcpp::export]]
SEXP test_qd_serialize(SEXP x) {
size_t len = 0;
unsigned char * buffer = c_qd_serialize(x, &len, 10, true, 4); // object, buffer length, compress_level, shuffle, nthreads
SEXP y = c_qd_deserialize(buffer, len, false, false, 4); // buffer, buffer length, use_alt_rep, validate_checksum, nthreads
c_qd_free(buffer); // must manually free buffer
return y;
}
/*** R
x <- runif(1e7)
stopifnot(test_qs_serialize(x) == x)
stopifnot(test_qd_serialize(x) == x)
*/