Skip to contents

This function allows you to write a dataset. By writing to more efficient binary storage formats, and by specifying relevant partitioning, you can make it much faster to read and query.


  format = c("parquet", "feather", "arrow", "ipc", "csv", "tsv", "txt", "text"),
  partitioning = dplyr::group_vars(dataset),
  basename_template = paste0("part-{i}.", as.character(format)),
  hive_style = TRUE,
  existing_data_behavior = c("overwrite", "error", "delete_matching"),
  max_partitions = 1024L,
  max_open_files = 900L,
  max_rows_per_file = 0L,
  min_rows_per_group = 0L,
  max_rows_per_group = bitwShiftL(1, 20),



Dataset, RecordBatch, Table, arrow_dplyr_query, or data.frame. If an arrow_dplyr_query, the query will be evaluated and the result will be written. This means that you can select(), filter(), mutate(), etc. to transform the data before it is written if you need to.


string path, URI, or SubTreeFileSystem referencing a directory to write to (directory will be created if it does not exist)


a string identifier of the file format. Default is to use "parquet" (see FileFormat)


Partitioning or a character vector of columns to use as partition keys (to be written as path segments). Default is to use the current group_by() columns.


string template for the names of files to be written. Must contain "{i}", which will be replaced with an autoincremented integer to generate basenames of datafiles. For example, "part-{i}.arrow" will yield "part-0.arrow", .... If not specified, it defaults to "part-{i}.<default extension>".


logical: write partition segments as Hive-style (key1=value1/key2=value2/file.ext) or as just bare values. Default is TRUE.


The behavior to use when there is already data in the destination directory. Must be one of "overwrite", "error", or "delete_matching".

  • "overwrite" (the default) then any new files created will overwrite existing files

  • "error" then the operation will fail if the destination directory is not empty

  • "delete_matching" then the writer will delete any existing partitions if data is going to be written to those partitions and will leave alone partitions which data is not written to.


maximum number of partitions any batch may be written into. Default is 1024L.


maximum number of files that can be left opened during a write operation. If greater than 0 then this will limit the maximum number of files that can be left open. If an attempt is made to open too many files then the least recently used file will be closed. If this setting is set too low you may end up fragmenting your data into many small files. The default is 900 which also allows some # of files to be open by the scanner before hitting the default Linux limit of 1024.


maximum number of rows per file. If greater than 0 then this will limit how many rows are placed in any single file. Default is 0L.


write the row groups to the disk when this number of rows have accumulated. Default is 0L.


maximum rows allowed in a single group and when this number of rows is exceeded, it is split and the next set of rows is written to the next group. This value must be set such that it is greater than min_rows_per_group. Default is 1024 * 1024.


additional format-specific arguments. For available Parquet options, see write_parquet(). The available Feather options are:

  • use_legacy_format logical: write data formatted so that Arrow libraries versions 0.14 and lower can read it. Default is FALSE. You can also enable this by setting the environment variable ARROW_PRE_0_15_IPC_FORMAT=1.

  • metadata_version: A string like "V5" or the equivalent integer indicating the Arrow IPC MetadataVersion. Default (NULL) will use the latest version, unless the environment variable ARROW_PRE_1_0_METADATA_VERSION=1, in which case it will be V4.

  • codec: A Codec which will be used to compress body buffers of written files. Default (NULL) will not compress body buffers.

  • null_fallback: character to be used in place of missing values (NA or NULL) when using Hive-style partitioning. See hive_partition().


The input dataset, invisibly


# You can write datasets partitioned by the values in a column (here: "cyl").
# This creates a structure of the form cyl=X/part-Z.parquet.
one_level_tree <- tempfile()
write_dataset(mtcars, one_level_tree, partitioning = "cyl")
list.files(one_level_tree, recursive = TRUE)
#> [1] "cyl=4/part-0.parquet" "cyl=6/part-0.parquet" "cyl=8/part-0.parquet"

# You can also partition by the values in multiple columns
# (here: "cyl" and "gear").
# This creates a structure of the form cyl=X/gear=Y/part-Z.parquet.
two_levels_tree <- tempfile()
write_dataset(mtcars, two_levels_tree, partitioning = c("cyl", "gear"))
list.files(two_levels_tree, recursive = TRUE)
#> [1] "cyl=4/gear=3/part-0.parquet" "cyl=4/gear=4/part-0.parquet"
#> [3] "cyl=4/gear=5/part-0.parquet" "cyl=6/gear=3/part-0.parquet"
#> [5] "cyl=6/gear=4/part-0.parquet" "cyl=6/gear=5/part-0.parquet"
#> [7] "cyl=8/gear=3/part-0.parquet" "cyl=8/gear=5/part-0.parquet"

# In the two previous examples we would have:
# X = {4,6,8}, the number of cylinders.
# Y = {3,4,5}, the number of forward gears.
# Z = {0,1,2}, the number of saved parts, starting from 0.

# You can obtain the same result as as the previous examples using arrow with
# a dplyr pipeline. This will be the same as two_levels_tree above, but the
# output directory will be different.
two_levels_tree_2 <- tempfile()
mtcars %>%
  group_by(cyl, gear) %>%
list.files(two_levels_tree_2, recursive = TRUE)
#> [1] "cyl=4/gear=3/part-0.parquet" "cyl=4/gear=4/part-0.parquet"
#> [3] "cyl=4/gear=5/part-0.parquet" "cyl=6/gear=3/part-0.parquet"
#> [5] "cyl=6/gear=4/part-0.parquet" "cyl=6/gear=5/part-0.parquet"
#> [7] "cyl=8/gear=3/part-0.parquet" "cyl=8/gear=5/part-0.parquet"

# And you can also turn off the Hive-style directory naming where the column
# name is included with the values by using `hive_style = FALSE`.

# Write a structure X/Y/part-Z.parquet.
two_levels_tree_no_hive <- tempfile()
mtcars %>%
  group_by(cyl, gear) %>%
  write_dataset(two_levels_tree_no_hive, hive_style = FALSE)
list.files(two_levels_tree_no_hive, recursive = TRUE)
#> [1] "4/3/part-0.parquet" "4/4/part-0.parquet" "4/5/part-0.parquet"
#> [4] "6/3/part-0.parquet" "6/4/part-0.parquet" "6/5/part-0.parquet"
#> [7] "8/3/part-0.parquet" "8/5/part-0.parquet"