- load a SAS dataset in parallel into a Spark table for further processing with the sparklyr package
- process in parallel the full SAS dataset with dplyr statements, instead of having to import the full SAS dataset in RAM (using the foreign/haven packages) and hence avoiding RAM problems of large imports
The following example reads in a file called iris.sas7bdat in a table called sas_example in Spark. Do try this with bigger data on your cluster and look at the help of the sparklyr package to connect to your Spark cluster.
library(sparklyr) library(spark.sas7bdat) mysasfile <- system.file("extdata", "iris.sas7bdat", package = "spark.sas7bdat") sc <- spark_connect(master = "local") x <- spark_read_sas(sc, path = mysasfile, table = "sas_example") x
The resulting pointer to a Spark table can be further used in dplyr statements
library(dplyr) x %>% group_by(Species) %>% summarise(count = n(), length = mean(Sepal_Length), width = mean(Sepal_Width))
Install the package from CRAN.
Or install this development version from github.
devtools::install_github("bnosac/spark.sas7bdat", build_vignettes = TRUE) vignette("spark_sas7bdat_examples", package = "spark.sas7bdat")
The package has been tested out with Spark version 2.0.1 and Hadoop 2.7.
library(sparklyr) spark_install(version = "2.0.1", hadoop_version = "2.7")
In order to compare the functionality to the read_sas function from the haven package, below we show a comparison on a small 5234557 rows x 2 columns SAS dataset with only numeric data. Processing is done on 8 cores. With the haven package you need to import the data in RAM, with the spark.sas7bdat package, you can immediately execute dplyr statements on top of the SAS dataset.
mysasfile <- "/home/bnosac/Desktop/testdata.sas7bdat" system.time(x <- spark_read_sas(sc, path = mysasfile, table = "testdata")) user system elapsed 0.008 0.000 0.051 system.time(x <- haven::read_sas(mysasfile)) user system elapsed 1.172 0.032 1.200
Support in big data and Spark analysis
Need support in big data and Spark analysis? Contact BNOSAC: http://www.bnosac.be