Skip to contents

Convert huge input file to parquet

Source: vignettes/aa-conversions.Rmd
aa-conversions.Rmd

With table_to_parquet()

For huge input files in SAS, SPSS and Stata formats, the parquetize package allows you to perform a clever conversion by using max_memory or max_rows in the table_to_parquet() function. The native behavior of this function (and all other functions in the package) is to load the entire table to be converted into R and then write it to disk (in a single file or a partitioned directory).

When handling very large files, the risk that frequently occurs is that the R session aborts because it cannot load the entire database into memory. This risk is even more present if you work locally on your computer and it can be limited if you work on remote servers.
table_to_parquet() offers this solution which answers a need expressed by parquetize users.

The idea is to split the very large table into “chunks” based on memory consumption of input data or on the number of rows in the table in order to be able to simultaneously :
- read a chunk of the very large database
- write this chunk in the floor file

Here are examples from the documentation using the iris table. There’s two ways to split output files :

  • by memory consumption
  • by number of lines

Spliting data by memory consumption

table_to_parquet can guess the number of lines to put in a file based on the memory consuption with the argument max_memory expressed in Mb.

Here we cut the 150 rows into chunks of roughly 5 Kb when a file is loaded as a tibble.
In this example we get 2 parquet files of 89 lines called iris1-89.parquet and iris90-150.parquet

table_to_parquet(
  path_to_file = system.file("examples", "iris.sas7bdat", package = "haven"),
  path_to_parquet = tempfile(),
  max_memory = 5 / 1024,
  encoding = "utf-8"
)
#> Reading data...
#> Writing file194e26f0e443-1-89.parquet...
#> Reading data...
#> Writing file194e26f0e443-90-150.parquet...
#> ✔ Data are available in parquet dataset under /tmp/Rtmp4DksHD/file194e26f0e443/
#> Writing file194e26f0e443-90-150.parquet...

In real life, you should use a max_memory in the Gb range, for example with a SAS file of 50 000 000 lines and using max_memory of 5000 Mb :

  table_to_parquet(
  path_to_file = "myhugefile.sas7bdat",
  path_to_parquet = tempdir(),
  max_memory = 5000,
  encoding = "utf-8"
)

Splitting data by number of lines

Tip: The number of lines that each chunk must contain must be supported by the RAM of your computer/server. Ideally, the number of chunks to be defined must be limited. It should be in tens and not hundreds to limit the number of intermediate files (see example below).

Here we cut the 150 rows into 3 chunks of 50 rows. In this example we get 3 parquet files of 50 lines called iris1-50.parquet, iris51-100.parquet and iris101-151.parquet

table_to_parquet(
  path_to_file = system.file("examples", "iris.sas7bdat", package = "haven"),
  path_to_parquet = tempfile(),
  max_rows = 50,
  encoding = "utf-8"
)

In real life, we can perform this kind of request with the parquetize API (for example with a SAS file of 50 000 000 lines and defining 25 chunks of 2 000 000 rows each) :

table_to_parquet(
  path_to_file = "myhugefile.sas7bdat",
  path_to_parquet = tempdir(),
  max_rows = 2000000,
  encoding = "utf-8"
)

Files myhugefile1-2000000.parquet, myhugefile2000001-4000000.parquet … will be created.

Function rbind_parquet()

If at the end of the conversion with table_to_parquet(), you want to reconstitute a unique initial table and if you have the computer resources (in RAM) to do so, you can use the helper function provided with the API of rbind_parquet().
This function allows to bind multiple parquet files by row.
Here’s an example without deleting initial files (delete_initial_files=FALSE) :

rbind_parquet(
  folder = tempfile(),
  output_name = "myhugefile",
  delete_initial_files = FALSE
)

The myhugefile.parquet file will be created from the myhugefile1-2000000.parquet, myhugefile2000001-4000000.parquet… files!

Alternatives to {parquetize}

Despite our best efforts, you may not be able to convert your very large database with {parquetize}.
In this case, one solution is probably to turn to duckdb, which offers undeniable advantages when it comes to conversion operations.