Vroom Benchmarks

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.

How it works

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.

Reading delimited files

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.

Taxi Trip Dataset

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 http://chriswhong.com/open-data/foil_nyc_taxi/. It is also hosted on archive.org.

The first table trip_fare_1.csv was converted to tsv and saved as trip_fare_1.tsv, It 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...

Taxi Benchmarks

The code used to run the taxi benchmarks is in bench/taxi-benchmark.R.

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, the benchmarks labeled vroom (full altrep) instead use Altrep vectors for all supported types.

The following operations are performed.

  • The data is read
  • print() - N.B. read.delim uses print(head(x, 10)) because printing the whole dataset takes > 10 minutes
  • head()
  • tail()
  • Sampling 100 random rows
  • Filtering for “UNK” payment, this is 6434 rows (0.0435% of total).
  • Aggregation of mean fare amount per payment type.
package read print head tail sample filter aggregate total
read.delim 1m 23.6s 6ms 1ms 1ms 1ms 336ms 745ms 1m 24.7s
readr 33.9s 115ms 1ms 1ms 2ms 214ms 899ms 35.1s
data.table 15.8s 16ms 1ms 1ms 1ms 133ms 397ms 16.4s
vroom base 3.8s 94ms 1ms 1ms 1ms 1.4s 7.2s 12.5s
vroom (full altrep) base 1.8s 130ms 1ms 1ms 1ms 1.4s 7.4s 10.7s
vroom (full altrep) dplyr 2.1s 156ms 1ms 1ms 3.2s 1.3s 2s 8.7s
vroom dplyr 3.8s 97ms 1ms 1ms 3ms 1.4s 2s 7.2s

(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

The code used to run the all numeric benchmarks is in bench/all_numeric-benchmark.R.

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 the vroom is multi-threaded and therefore is quicker than readr and read.delim.

package read print head tail sample filter aggregate total
read.delim 2m 0.5s 9ms 1ms 1ms 2ms 7.6s 54ms 2m 8.2s
readr 5.5s 102ms 1ms 1ms 3ms 11ms 44ms 5.7s
vroom base 1.6s 119ms 1ms 1ms 3ms 82ms 58ms 1.9s
vroom dplyr 1.5s 119ms 1ms 1ms 3ms 13ms 41ms 1.7s
vroom (full altrep) dplyr 354ms 112ms 1ms 1ms 1.1s 63ms 47ms 1.7s
vroom (full altrep) base 488ms 157ms 1ms 1ms 3ms 64ms 253ms 963ms
data.table 513ms 18ms 1ms 1ms 4ms 6ms 59ms 598ms

All character data

The code used to run the all character benchmarks is in bench/all_character-benchmark.R.

All character data is a best case scenario for vroom, as none of the data needs to be read initially.

package read print head tail sample filter aggregate total
read.delim 1m 41.9s 9ms 1ms 1ms 3ms 23ms 338ms 1m 42.3s
readr 1m 1.2s 112ms 1ms 1ms 5ms 17ms 369ms 1m 1.7s
data.table 44.3s 23ms 1ms 1ms 10ms 27ms 268ms 44.6s
vroom (full altrep) base 521ms 146ms 1ms 1ms 3ms 190ms 2s 2.9s
vroom base 506ms 131ms 1ms 1ms 3ms 142ms 2s 2.8s
vroom dplyr 387ms 162ms 1ms 1ms 5ms 173ms 1.3s 2s
vroom (full altrep) dplyr 429ms 105ms 1ms 1ms 4ms 179ms 1.3s 2s

Writing delimited files

The code used to run the taxi writing benchmarks is at bench/taxi_writing-benchmark.R.

The benchmarks write out the taxi trip dataset in three different ways.

Note the current CRAN version of data.table (1.12.2) does not support writing to compressed files

package uncompressed gzip multithreaded gzip
data.table 5.6s NA NA
write.delim 1m 36.4s 3m 19.6s 1m 29.3s
readr 1m 10s 2m 18.8s 1m 4.5s
vroom 6.5s 1m 19.5s 19s

Session and package information

The development version of dplyr was used in the benchmarks, as contains a fix for a performance issue with Altrep objects.

package version date source
base 3.5.3 2019-03-13 local
data.table 1.12.2 2019-04-07 CRAN (R 3.5.2)
dplyr 0.8.0.9012 2019-04-29 local
readr 1.3.1 2018-12-21 CRAN (R 3.5.0)
vroom 1.0.1 2019-05-14 local