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 by_chunk=TRUE 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 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

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 is an example from the documentation using the iris table. 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_table = system.file("examples", "iris.sas7bdat", package = "haven"),
path_to_parquet = tempfile(),
by_chunk = TRUE,
chunk_size = 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_table = "myhugefile.sas7bdat",
path_to_parquet = tempdir(),
by_chunk = TRUE,
chunk_size = 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. :tada: