knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) library(ggplot2) library(forcats) library(dplyr) library(tidyr) library(fs) pretty_sec <- function(x) { x[!is.na(x)] <- prettyunits::pretty_sec(x[!is.na(x)]) x } pretty_lgl <- function(x) { case_when( x == TRUE ~ "TRUE", x == FALSE ~ "FALSE", TRUE ~ "" ) } read_benchmark <- function(file, desc) { vroom::vroom(file, col_types = c("ccccddddd")) %>% filter(op != "setup") %>% mutate( altrep = case_when( grepl("^vroom_no_altrep", reading_package) ~ FALSE, grepl("^vroom", reading_package) ~ TRUE, TRUE ~ NA ), reading_package = case_when( grepl("^vroom", reading_package) ~ "vroom", TRUE ~ reading_package ), label = fct_reorder( glue::glue("{reading_package}{altrep}\n{manip_package}", altrep = ifelse(is.na(altrep), "", glue::glue("(altrep = {altrep})")) ), case_when(type == "real" ~ time, TRUE ~ 0), sum), op = factor(op, desc) ) } generate_subtitle <- function(data) { rows <- scales::comma(data$rows[[1]]) cols <- scales::comma(data$cols[[1]]) size <- fs_bytes(data$size[[1]]) glue::glue("{rows} x {cols} - {size}B") } plot_benchmark <- function(data, title) { subtitle <- generate_subtitle(data) data <- data %>% filter(reading_package != "read.delim", type == "real") p1 <- data %>% ggplot() + geom_bar(aes(x = label, y = time, fill = op, group = label), stat = "identity") + scale_fill_brewer(type = "qual", palette = "Set2") + scale_y_continuous(labels = scales::number_format(suffix = "s")) + coord_flip() + labs(title = title, subtitle = subtitle, x = NULL, y = NULL, fill = NULL) + theme(legend.position = "bottom") p2 <- data %>% group_by(label) %>% summarise(max_memory = max(max_memory)) %>% ggplot() + geom_bar(aes(x = label, y = max_memory / (1024 * 1024)), stat = "identity") + scale_y_continuous(labels = scales::number_format(suffix = "Mb")) + coord_flip() + labs(title = "Maximum memory usage", x = NULL, y = NULL) + theme(axis.text.y = element_blank(), axis.ticks.y = element_blank()) library(patchwork) p1 + p2 + plot_layout(widths = c(2, 1)) } make_table <- function(data) { data %>% filter(type == "real") %>% select(-label, -size, -type, -rows, -cols) %>% spread(op, time) %>% mutate( total = read + print + head + tail + sample + filter + aggregate, max_memory = as.character(bench::as_bench_bytes(max_memory)) ) %>% arrange(desc(total)) %>% mutate_if(is.numeric, pretty_sec) %>% mutate_if(is.logical, pretty_lgl) %>% select(reading_package, manip_package, altrep, max_memory, everything()) %>% rename( "reading\npackage" = reading_package, "manipulating\npackage" = manip_package, memory = max_memory ) %>% knitr::kable(digits = 2, align = "r", format = "html") } desc <- c("setup", "read", "print", "head", "tail", "sample", "filter", "aggregate")
vroom is a new approach to reading delimited and fixed width data into R.
It stems from the observation that when parsing files reading data from disk and finding the delimiters is generally not the main bottle neck. Instead (re)-allocating memory and parsing the values into R data types (particularly for characters) takes the bulk of the time.
Therefore you can obtain very rapid input by first performing a fast indexing step and then using the Altrep framework available in R versions 3.5+ to access the values in a lazy / delayed fashion.
The initial reading of the file simply records the locations of each individual record, the actual values are not read into R. Altrep vectors are created for each column in the data which hold a pointer to the index and the memory mapped file. When these vectors are indexed the value is read from the memory mapping.
This means initial reading is extremely fast, in the real world dataset below
it is ~ 1/4 the time of the multi-threaded data.table::fread()
. Sampling
operations are likewise extremely fast, as only the data actually included in
the sample is read. This means things like the tibble print method, calling
head()
, tail()
x[sample(), ]
etc. have very low overhead. Filtering also
can be fast, only the columns included in the filter selection have to be fully
read and only the data in the filtered rows needs to be read from the remaining
columns. Grouped aggregations likewise only need to read the grouping
variables and the variables aggregated.
Once a particular vector is fully materialized the speed for all subsequent operations should be identical to a normal R vector.
This approach potentially also allows you to work with data that is larger than memory. As long as you are careful to avoid materializing the entire dataset at once it can be efficiently queried and subset.
The following benchmarks all measure reading delimited files of various sizes and data types. Because vroom delays reading the benchmarks also do some manipulation of the data afterwards to try and provide a more realistic performance comparison.
Because the read.delim
results are so much slower than the others they are
excluded from the plots, but are retained in the tables.
This real world dataset is from Freedom of Information Law (FOIL) Taxi Trip Data from the NYC Taxi and Limousine Commission 2013, originally posted at https://chriswhong.com/open-data/foil_nyc_taxi/. It is also hosted on archive.org.
The first table trip_fare_1.csv is 1.55G in size.
#> Observations: 14,776,615 #> Variables: 11 #> $ medallion <chr> "89D227B655E5C82AECF13C3F540D4CF4", "0BD7C8F5B... #> $ hack_license <chr> "BA96DE419E711691B9445D6A6307C170", "9FD8F69F0... #> $ vendor_id <chr> "CMT", "CMT", "CMT", "CMT", "CMT", "CMT", "CMT... #> $ pickup_datetime <chr> "2013-01-01 15:11:48", "2013-01-06 00:18:35", ... #> $ payment_type <chr> "CSH", "CSH", "CSH", "CSH", "CSH", "CSH", "CSH... #> $ fare_amount <dbl> 6.5, 6.0, 5.5, 5.0, 9.5, 9.5, 6.0, 34.0, 5.5, ... #> $ surcharge <dbl> 0.0, 0.5, 1.0, 0.5, 0.5, 0.0, 0.0, 0.0, 1.0, 0... #> $ mta_tax <dbl> 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0... #> $ tip_amount <int> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0... #> $ tolls_amount <dbl> 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 4.8, 0.0, 0... #> $ total_amount <dbl> 7.0, 7.0, 7.0, 6.0, 10.5, 10.0, 6.5, 39.3, 7.0...
code: bench/taxi
All benchmarks were run on a Amazon EC2 m5.4xlarge instance with 16 vCPUs and an EBS volume type.
The benchmarks labeled vroom_base
uses vroom
with base functions for
manipulation. vroom_dplyr
uses vroom
to read the file and dplyr functions to
manipulate. data.table
uses fread()
to read the file and data.table
functions to manipulate and readr
uses readr
to read the file and dplyr
to
manipulate. By default vroom only uses Altrep for character vectors, these are
labeled vroom(altrep: normal)
. The benchmarks labeled vroom(altrep: full)
instead use Altrep vectors for all supported types and vroom(altrep: none)
disable Altrep entirely.
The following operations are performed.
print()
- N.B. read.delim uses print(head(x, 10))
because printing the whole dataset takes > 10 minuteshead()
tail()
taxi <- read_benchmark(path_package("vroom", "bench", "taxi.tsv"), desc) plot_benchmark(taxi, "Time to analyze taxi trip data") make_table(taxi)
(N.B. Rcpp used in the dplyr implementation
fully materializes all the Altrep numeric vectors when using filter()
or sample_n()
,
which is why the first of these cases have additional overhead when using full Altrep.).
All numeric data is really a worst case scenario for vroom. The index takes about as much memory as the parsed data. Also because parsing doubles can be done quickly in parallel and text representations of doubles are only ~25 characters at most there isn't a great deal of savings for delayed parsing.
For these reasons (and because the data.table implementation is very fast) vroom is a bit slower than fread for pure numeric data.
However because vroom is multi-threaded it is a bit quicker than readr and read.delim for this type of data.
code: bench/all_numeric-long
all_num <- read_benchmark(path_package("vroom", "bench", "all_numeric-long.tsv"), desc) plot_benchmark(all_num, "Time to analyze long all numeric data") make_table(all_num)
code: bench/all_numeric-wide
all_num_wide <- read_benchmark(path_package("bench", "all_numeric-wide.tsv", package = "vroom"), desc) plot_benchmark(all_num_wide, "Time to analyze wide all numeric data") make_table(all_num_wide)
code: bench/all_character-long
All character data is a best case scenario for vroom when using Altrep, as it takes full advantage of the lazy reading.
all_chr <- read_benchmark(path_package("vroom", "bench", "all_character-long.tsv"), desc) plot_benchmark(all_chr, "Time to analyze long all character data") make_table(all_chr)
code: bench/all_character-wide
all_chr_wide <- read_benchmark(path_package("vroom", "bench", "all_character-wide.tsv"), desc) plot_benchmark(all_chr_wide, "Time to analyze wide all character data") make_table(all_chr_wide)
code: bench/taxi_multiple
mult <- read_benchmark(path_package("vroom", "bench", "taxi_multiple.tsv"), desc)
The benchmark reads all 12 files in the taxi trip fare data, totaling r scales::comma(mult$rows[[1]])
rows
and r mult$cols[[1]]
columns for a total file size of r format(fs_bytes(mult$size[[1]]))
.
plot_benchmark(mult, "Time to analyze multiple file data") make_table(mult)
fwf <- read_benchmark(path_package("vroom", "bench", "fwf.tsv"), desc)
This fixed width dataset contains individual records of the characteristics of a 5 percent sample of people and housing units from the year 2000 and is freely available at https://web.archive.org/web/20150908055439/https://www2.census.gov/census_2000/datasets/PUMS/FivePercent/California/all_California.zip. The data is split into files by state, and the state of California was used in this benchmark.
The data totals r scales::comma(fwf$rows[[1]])
rows
and r fwf$cols[[1]]
columns with a total file size of r format(fs_bytes(fwf$size[[1]]))
.
code: bench/fwf
plot_benchmark(fwf, "Time to analyze fixed width data") make_table(fwf)
code: bench/taxi_writing
The benchmarks write out the taxi trip dataset in a few different ways.
gzfile()
(readr and vroom do this automatically for files ending in .gz
)pipe()
connection to pigz for the rest).taxi_writing <- read_benchmark(path_package("vroom", "bench", "taxi_writing.tsv"), c("setup", "writing")) %>% rename( package = reading_package, compression = manip_package ) %>% mutate( package = factor(package, c("base", "readr", "data.table", "vroom")), compression = factor(compression, rev(c("gzip", "multithreaded_gzip", "zstandard", "uncompressed"))) ) %>% filter(type == "real") subtitle <- generate_subtitle(taxi_writing) taxi_writing %>% ggplot(aes(x = compression, y = time, fill = package)) + geom_bar(stat = "identity", position = position_dodge2(reverse = TRUE, padding = .05)) + scale_fill_brewer(type = "qual", palette = "Set2") + scale_y_continuous(labels = scales::number_format(suffix = "s")) + theme(legend.position = "bottom") + coord_flip() + labs(title = "Writing taxi trip data", subtitle = subtitle, x = NULL, y = NULL, fill = NULL) taxi_writing %>% select(-size, -op, -rows, -cols, -type, -altrep, -label, -max_memory) %>% mutate_if(is.numeric, pretty_sec) %>% pivot_wider(names_from = package, values_from = time) %>% arrange(desc(compression)) %>% knitr::kable(digits = 2, align = "r", format = "html")
si <- vroom::vroom(path_package("vroom", "bench", "session_info.tsv")) class(si) <- c("packages_info", "data.frame") select(as.data.frame(si), package, version = ondiskversion, date, source) %>% knitr::kable()
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.