SparkR (R on Spark)

Overview

SparkR is an R package that provides a light-weight frontend to use Apache Spark from R. In Spark 3.0.0-preview, SparkR provides a distributed data frame implementation that supports operations like selection, filtering, aggregation etc. (similar to R data frames, dplyr) but on large datasets. SparkR also supports distributed machine learning using MLlib.

SparkDataFrame

A SparkDataFrame is a distributed collection of data organized into named columns. It is conceptually equivalent to a table in a relational database or a data frame in R, but with richer optimizations under the hood. SparkDataFrames can be constructed from a wide array of sources such as: structured data files, tables in Hive, external databases, or existing local R data frames.

All of the examples on this page use sample data included in R or the Spark distribution and can be run using the ./bin/sparkR shell.

Starting Up: SparkSession

The entry point into SparkR is the SparkSession which connects your R program to a Spark cluster. You can create a SparkSession using sparkR.session and pass in options such as the application name, any spark packages depended on, etc. Further, you can also work with SparkDataFrames via SparkSession. If you are working from the sparkR shell, the SparkSession should already be created for you, and you would not need to call sparkR.session.

sparkR.session()

Starting Up from RStudio

You can also start SparkR from RStudio. You can connect your R program to a Spark cluster from RStudio, R shell, Rscript or other R IDEs. To start, make sure SPARK_HOME is set in environment (you can check Sys.getenv), load the SparkR package, and call sparkR.session as below. It will check for the Spark installation, and, if not found, it will be downloaded and cached automatically. Alternatively, you can also run install.spark manually.

In addition to calling sparkR.session, you could also specify certain Spark driver properties. Normally these Application properties and Runtime Environment cannot be set programmatically, as the driver JVM process would have been started, in this case SparkR takes care of this for you. To set them, pass them as you would other configuration properties in the sparkConfig argument to sparkR.session().

if (nchar(Sys.getenv("SPARK_HOME")) < 1) {
  Sys.setenv(SPARK_HOME = "/home/spark")
}
library(SparkR, lib.loc = c(file.path(Sys.getenv("SPARK_HOME"), "R", "lib")))
sparkR.session(master = "local[*]", sparkConfig = list(spark.driver.memory = "2g"))

The following Spark driver properties can be set in sparkConfig with sparkR.session from RStudio:

Property NameProperty groupspark-submit equivalent
spark.master Application Properties --master
spark.kerberos.keytab Application Properties --keytab
spark.kerberos.principal Application Properties --principal
spark.driver.memory Application Properties --driver-memory
spark.driver.extraClassPath Runtime Environment --driver-class-path
spark.driver.extraJavaOptions Runtime Environment --driver-java-options
spark.driver.extraLibraryPath Runtime Environment --driver-library-path

Creating SparkDataFrames

With a SparkSession, applications can create SparkDataFrames from a local R data frame, from a Hive table, or from other data sources.

From local data frames

The simplest way to create a data frame is to convert a local R data frame into a SparkDataFrame. Specifically, we can use as.DataFrame or createDataFrame and pass in the local R data frame to create a SparkDataFrame. As an example, the following creates a SparkDataFrame based using the faithful dataset from R.

df <- as.DataFrame(faithful)

# Displays the first part of the SparkDataFrame
head(df)
##  eruptions waiting
##1     3.600      79
##2     1.800      54
##3     3.333      74

From Data Sources

SparkR supports operating on a variety of data sources through the SparkDataFrame interface. This section describes the general methods for loading and saving data using Data Sources. You can check the Spark SQL programming guide for more specific options that are available for the built-in data sources.

The general method for creating SparkDataFrames from data sources is read.df. This method takes in the path for the file to load and the type of data source, and the currently active SparkSession will be used automatically. SparkR supports reading JSON, CSV and Parquet files natively, and through packages available from sources like Third Party Projects, you can find data source connectors for popular file formats like Avro. These packages can either be added by specifying --packages with spark-submit or sparkR commands, or if initializing SparkSession with sparkPackages parameter when in an interactive R shell or from RStudio.

sparkR.session(sparkPackages = "org.apache.spark:spark-avro_2.12:3.0.0-preview")

We can see how to use data sources using an example JSON input file. Note that the file that is used here is not a typical JSON file. Each line in the file must contain a separate, self-contained valid JSON object. For more information, please see JSON Lines text format, also called newline-delimited JSON. As a consequence, a regular multi-line JSON file will most often fail.

people <- read.df("./examples/src/main/resources/people.json", "json")
head(people)
##  age    name
##1  NA Michael
##2  30    Andy
##3  19  Justin

# SparkR automatically infers the schema from the JSON file
printSchema(people)
# root
#  |-- age: long (nullable = true)
#  |-- name: string (nullable = true)

# Similarly, multiple files can be read with read.json
people <- read.json(c("./examples/src/main/resources/people.json", "./examples/src/main/resources/people2.json"))

The data sources API natively supports CSV formatted input files. For more information please refer to SparkR read.df API documentation.

df <- read.df(csvPath, "csv", header = "true", inferSchema = "true", na.strings = "NA")

The data sources API can also be used to save out SparkDataFrames into multiple file formats. For example, we can save the SparkDataFrame from the previous example to a Parquet file using write.df.

write.df(people, path = "people.parquet", source = "parquet", mode = "overwrite")

From Hive tables

You can also create SparkDataFrames from Hive tables. To do this we will need to create a SparkSession with Hive support which can access tables in the Hive MetaStore. Note that Spark should have been built with Hive support and more details can be found in the SQL programming guide. In SparkR, by default it will attempt to create a SparkSession with Hive support enabled (enableHiveSupport = TRUE).

sparkR.session()

sql("CREATE TABLE IF NOT EXISTS src (key INT, value STRING)")
sql("LOAD DATA LOCAL INPATH 'examples/src/main/resources/kv1.txt' INTO TABLE src")

# Queries can be expressed in HiveQL.
results <- sql("FROM src SELECT key, value")

# results is now a SparkDataFrame
head(results)
##  key   value
## 1 238 val_238
## 2  86  val_86
## 3 311 val_311

SparkDataFrame Operations

SparkDataFrames support a number of functions to do structured data processing. Here we include some basic examples and a complete list can be found in the API docs:

Selecting rows, columns

# Create the SparkDataFrame
df <- as.DataFrame(faithful)

# Get basic information about the SparkDataFrame
df
## SparkDataFrame[eruptions:double, waiting:double]

# Select only the "eruptions" column
head(select(df, df$eruptions))
##  eruptions
##1     3.600
##2     1.800
##3     3.333

# You can also pass in column name as strings
head(select(df, "eruptions"))

# Filter the SparkDataFrame to only retain rows with wait times shorter than 50 mins
head(filter(df, df$waiting < 50))
##  eruptions waiting
##1     1.750      47
##2     1.750      47
##3     1.867      48

Grouping, Aggregation

SparkR data frames support a number of commonly used functions to aggregate data after grouping. For example, we can compute a histogram of the waiting time in the faithful dataset as shown below

# We use the `n` operator to count the number of times each waiting time appears
head(summarize(groupBy(df, df$waiting), count = n(df$waiting)))
##  waiting count
##1      70     4
##2      67     1
##3      69     2

# We can also sort the output from the aggregation to get the most common waiting times
waiting_counts <- summarize(groupBy(df, df$waiting), count = n(df$waiting))
head(arrange(waiting_counts, desc(waiting_counts$count)))
##   waiting count
##1      78    15
##2      83    14
##3      81    13

In addition to standard aggregations, SparkR supports OLAP cube operators cube:

head(agg(cube(df, "cyl", "disp", "gear"), avg(df$mpg)))
##  cyl  disp gear avg(mpg)
##1  NA 140.8    4     22.8
##2   4  75.7    4     30.4
##3   8 400.0    3     19.2
##4   8 318.0    3     15.5
##5  NA 351.0   NA     15.8
##6  NA 275.8   NA     16.3

and rollup:

head(agg(rollup(df, "cyl", "disp", "gear"), avg(df$mpg)))
##  cyl  disp gear avg(mpg)
##1   4  75.7    4     30.4
##2   8 400.0    3     19.2
##3   8 318.0    3     15.5
##4   4  78.7   NA     32.4
##5   8 304.0    3     15.2
##6   4  79.0   NA     27.3

Operating on Columns

SparkR also provides a number of functions that can be directly applied to columns for data processing and during aggregation. The example below shows the use of basic arithmetic functions.

# Convert waiting time from hours to seconds.
# Note that we can assign this to a new column in the same SparkDataFrame
df$waiting_secs <- df$waiting * 60
head(df)
##  eruptions waiting waiting_secs
##1     3.600      79         4740
##2     1.800      54         3240
##3     3.333      74         4440

Applying User-Defined Function

In SparkR, we support several kinds of User-Defined Functions:

Run a given function on a large dataset using dapply or dapplyCollect

dapply

Apply a function to each partition of a SparkDataFrame. The function to be applied to each partition of the SparkDataFrame and should have only one parameter, to which a data.frame corresponds to each partition will be passed. The output of function should be a data.frame. Schema specifies the row format of the resulting a SparkDataFrame. It must match to data types of returned value.

# Convert waiting time from hours to seconds.
# Note that we can apply UDF to DataFrame.
schema <- structType(structField("eruptions", "double"), structField("waiting", "double"),
                     structField("waiting_secs", "double"))
df1 <- dapply(df, function(x) { x <- cbind(x, x$waiting * 60) }, schema)
head(collect(df1))
##  eruptions waiting waiting_secs
##1     3.600      79         4740
##2     1.800      54         3240
##3     3.333      74         4440
##4     2.283      62         3720
##5     4.533      85         5100
##6     2.883      55         3300
dapplyCollect

Like dapply, apply a function to each partition of a SparkDataFrame and collect the result back. The output of function should be a data.frame. But, Schema is not required to be passed. Note that dapplyCollect can fail if the output of UDF run on all the partition cannot be pulled to the driver and fit in driver memory.

# Convert waiting time from hours to seconds.
# Note that we can apply UDF to DataFrame and return a R's data.frame
ldf <- dapplyCollect(
         df,
         function(x) {
           x <- cbind(x, "waiting_secs" = x$waiting * 60)
         })
head(ldf, 3)
##  eruptions waiting waiting_secs
##1     3.600      79         4740
##2     1.800      54         3240
##3     3.333      74         4440

Run a given function on a large dataset grouping by input column(s) and using gapply or gapplyCollect

gapply

Apply a function to each group of a SparkDataFrame. The function is to be applied to each group of the SparkDataFrame and should have only two parameters: grouping key and R data.frame corresponding to that key. The groups are chosen from SparkDataFrames column(s). The output of function should be a data.frame. Schema specifies the row format of the resulting SparkDataFrame. It must represent R function’s output schema on the basis of Spark data types. The column names of the returned data.frame are set by user.

# Determine six waiting times with the largest eruption time in minutes.
schema <- structType(structField("waiting", "double"), structField("max_eruption", "double"))
result <- gapply(
    df,
    "waiting",
    function(key, x) {
        y <- data.frame(key, max(x$eruptions))
    },
    schema)
head(collect(arrange(result, "max_eruption", decreasing = TRUE)))

##    waiting   max_eruption
##1      64       5.100
##2      69       5.067
##3      71       5.033
##4      87       5.000
##5      63       4.933
##6      89       4.900
gapplyCollect

Like gapply, applies a function to each partition of a SparkDataFrame and collect the result back to R data.frame. The output of the function should be a data.frame. But, the schema is not required to be passed. Note that gapplyCollect can fail if the output of UDF run on all the partition cannot be pulled to the driver and fit in driver memory.

# Determine six waiting times with the largest eruption time in minutes.
result <- gapplyCollect(
    df,
    "waiting",
    function(key, x) {
        y <- data.frame(key, max(x$eruptions))
        colnames(y) <- c("waiting", "max_eruption")
        y
    })
head(result[order(result$max_eruption, decreasing = TRUE), ])

##    waiting   max_eruption
##1      64       5.100
##2      69       5.067
##3      71       5.033
##4      87       5.000
##5      63       4.933
##6      89       4.900

Run local R functions distributed using spark.lapply

spark.lapply

Similar to lapply in native R, spark.lapply runs a function over a list of elements and distributes the computations with Spark. Applies a function in a manner that is similar to doParallel or lapply to elements of a list. The results of all the computations should fit in a single machine. If that is not the case they can do something like df <- createDataFrame(list) and then use dapply

# Perform distributed training of multiple models with spark.lapply. Here, we pass
# a read-only list of arguments which specifies family the generalized linear model should be.
families <- c("gaussian", "poisson")
train <- function(family) {
  model <- glm(Sepal.Length ~ Sepal.Width + Species, iris, family = family)
  summary(model)
}
# Return a list of model's summaries
model.summaries <- spark.lapply(families, train)

# Print the summary of each model
print(model.summaries)

Eager execution

If eager execution is enabled, the data will be returned to R client immediately when the SparkDataFrame is created. By default, eager execution is not enabled and can be enabled by setting the configuration property spark.sql.repl.eagerEval.enabled to true when the SparkSession is started up.

Maximum number of rows and maximum number of characters per column of data to display can be controlled by spark.sql.repl.eagerEval.maxNumRows and spark.sql.repl.eagerEval.truncate configuration properties, respectively. These properties are only effective when eager execution is enabled. If these properties are not set explicitly, by default, data up to 20 rows and up to 20 characters per column will be showed.

# Start up spark session with eager execution enabled
sparkR.session(master = "local[*]",
               sparkConfig = list(spark.sql.repl.eagerEval.enabled = "true",
                                  spark.sql.repl.eagerEval.maxNumRows = as.integer(10)))

# Create a grouped and sorted SparkDataFrame
df <- createDataFrame(faithful)
df2 <- arrange(summarize(groupBy(df, df$waiting), count = n(df$waiting)), "waiting")

# Similar to R data.frame, displays the data returned, instead of SparkDataFrame class string
df2

##+-------+-----+
##|waiting|count|
##+-------+-----+
##|   43.0|    1|
##|   45.0|    3|
##|   46.0|    5|
##|   47.0|    4|
##|   48.0|    3|
##|   49.0|    5|
##|   50.0|    5|
##|   51.0|    6|
##|   52.0|    5|
##|   53.0|    7|
##+-------+-----+
##only showing top 10 rows

Note that to enable eager execution in sparkR shell, add spark.sql.repl.eagerEval.enabled=true configuration property to the --conf option.

Running SQL Queries from SparkR

A SparkDataFrame can also be registered as a temporary view in Spark SQL and that allows you to run SQL queries over its data. The sql function enables applications to run SQL queries programmatically and returns the result as a SparkDataFrame.

# Load a JSON file
people <- read.df("./examples/src/main/resources/people.json", "json")

# Register this SparkDataFrame as a temporary view.
createOrReplaceTempView(people, "people")

# SQL statements can be run by using the sql method
teenagers <- sql("SELECT name FROM people WHERE age >= 13 AND age <= 19")
head(teenagers)
##    name
##1 Justin

Machine Learning

Algorithms

SparkR supports the following machine learning algorithms currently:

Classification

Regression

Tree

Clustering

Collaborative Filtering

Frequent Pattern Mining

Statistics

Under the hood, SparkR uses MLlib to train the model. Please refer to the corresponding section of MLlib user guide for example code. Users can call summary to print a summary of the fitted model, predict to make predictions on new data, and write.ml/read.ml to save/load fitted models. SparkR supports a subset of the available R formula operators for model fitting, including ‘~’, ‘.’, ‘:’, ‘+’, and ‘-‘.

Model persistence

The following example shows how to save/load a MLlib model by SparkR.

training <- read.df("data/mllib/sample_multiclass_classification_data.txt", source = "libsvm")
# Fit a generalized linear model of family "gaussian" with spark.glm
df_list <- randomSplit(training, c(7,3), 2)
gaussianDF <- df_list[[1]]
gaussianTestDF <- df_list[[2]]
gaussianGLM <- spark.glm(gaussianDF, label ~ features, family = "gaussian")

# Save and then load a fitted MLlib model
modelPath <- tempfile(pattern = "ml", fileext = ".tmp")
write.ml(gaussianGLM, modelPath)
gaussianGLM2 <- read.ml(modelPath)

# Check model summary
summary(gaussianGLM2)

# Check model prediction
gaussianPredictions <- predict(gaussianGLM2, gaussianTestDF)
head(gaussianPredictions)

unlink(modelPath)
Find full example code at "examples/src/main/r/ml/ml.R" in the Spark repo.

Data type mapping between R and Spark

RSpark
byte byte
integer integer
float float
double double
numeric double
character string
string string
binary binary
raw binary
logical boolean
POSIXct timestamp
POSIXlt timestamp
Date date
array array
list array
env map

Structured Streaming

SparkR supports the Structured Streaming API. Structured Streaming is a scalable and fault-tolerant stream processing engine built on the Spark SQL engine. For more information see the R API on the Structured Streaming Programming Guide

Apache Arrow in SparkR

Apache Arrow is an in-memory columnar data format that is used in Spark to efficiently transfer data between JVM and R processes. See also PySpark optimization done, PySpark Usage Guide for Pandas with Apache Arrow. This guide targets to explain how to use Arrow optimization in SparkR with some key points.

Ensure Arrow Installed

Arrow R library is available on CRAN as of ARROW-3204. It can be installed as below.

Rscript -e 'install.packages("arrow", repos="https://cloud.r-project.org/")'

If you need to install old versions, it should be installed directly from Github. You can use remotes::install_github as below.

Rscript -e 'remotes::install_github("apache/[email protected]", subdir = "r")'

apache-arrow-0.12.1 is a version tag that can be checked in Arrow at Github. You must ensure that Arrow R package is installed and available on all cluster nodes. The current supported minimum version is 0.12.1; however, this might change between the minor releases since Arrow optimization in SparkR is experimental.

Enabling for Conversion to/from R DataFrame, dapply and gapply

Arrow optimization is available when converting a Spark DataFrame to an R DataFrame using the call collect(spark_df), when creating a Spark DataFrame from an R DataFrame with createDataFrame(r_df), when applying an R native function to each partition via dapply(...) and when applying an R native function to grouped data via gapply(...). To use Arrow when executing these calls, users need to first set the Spark configuration ‘spark.sql.execution.arrow.sparkr.enabled’ to ‘true’. This is disabled by default.

In addition, optimizations enabled by ‘spark.sql.execution.arrow.sparkr.enabled’ could fallback automatically to non-Arrow optimization implementation if an error occurs before the actual computation within Spark during converting a Spark DataFrame to/from an R DataFrame.

# Start up spark session with Arrow optimization enabled
sparkR.session(master = "local[*]",
               sparkConfig = list(spark.sql.execution.arrow.sparkr.enabled = "true"))

# Converts Spark DataFrame from an R DataFrame
spark_df <- createDataFrame(mtcars)

# Converts Spark DataFrame to an R DataFrame
collect(spark_df)

# Apply an R native function to each partition.
collect(dapply(spark_df, function(rdf) { data.frame(rdf$gear + 1) }, structType("gear double")))

# Apply an R native function to grouped data.
collect(gapply(spark_df,
               "gear",
               function(key, group) {
                 data.frame(gear = key[[1]], disp = mean(group$disp) > group$disp)
               },
               structType("gear double, disp boolean")))

Using the above optimizations with Arrow will produce the same results as when Arrow is not enabled. Note that even with Arrow, collect(spark_df) results in the collection of all records in the DataFrame to the driver program and should be done on a small subset of the data.

Supported SQL Types

Currently, all Spark SQL data types are supported by Arrow-based conversion except FloatType, BinaryType, ArrayType, StructType and MapType.

R Function Name Conflicts

When loading and attaching a new package in R, it is possible to have a name conflict, where a function is masking another function.

The following functions are masked by the SparkR package:

Masked functionHow to Access
cov in package:stats
stats::cov(x, y = NULL, use = "everything",
           method = c("pearson", "kendall", "spearman"))
filter in package:stats
stats::filter(x, filter, method = c("convolution", "recursive"),
              sides = 2, circular = FALSE, init)
sample in package:base base::sample(x, size, replace = FALSE, prob = NULL)

Since part of SparkR is modeled on the dplyr package, certain functions in SparkR share the same names with those in dplyr. Depending on the load order of the two packages, some functions from the package loaded first are masked by those in the package loaded after. In such case, prefix such calls with the package name, for instance, SparkR::cume_dist(x) or dplyr::cume_dist(x).

You can inspect the search path in R with search()

Migration Guide

The migration guide is now archived on this page.