Fun Quiz on SparkR: How to Create Data Frame without Spark-CSV Package :)

1. Introduction

We are able to write Spark applications that create directly a data frame from a csv file by using the spark-csv package. And so, the package is often used.

Mahler's Symphony No. 5 will be performed by Eliahu Inbal and Konzerthausorchester Berlin in Tokyo tonight. I arrived at the hall an hour before that the classical concert started. So, I considered a few interesting quiz about creating data frames from csv files without above package :)

In this article, Spark's version is 1.6.3.

2. Fun Quiz

In this chapter I note the source code using the spark-csv package and the few funny quiz :)
Before starting the quiz game, we need to prepare the following csv file.

・ test_data.csv

20170228,scheme1,BermudanCallableSwap,JPY,21261339422
20170228,scheme2,BermudanCallableSwap,JPY,22759109989
20170228,scheme3,BermudanCallableSwap,JPY,21405741891
...

2.1. Use Spark-CSV Package

First, I note the program using the spark-csv package. There are two important points. One is setting 'SPARKR_SUBMIT_ARGS' to use spark-csv package, and the other is to load a csv file as data frame directly. 

How to set 'SPARKR_SUBMIT_ARGS' and load a csv file as data frame are as follows.

・ set 'SPARKR_SUBMIT_ARGS' (com.databricks:spark-csv)

Sys.setenv('SPARKR_SUBMIT_ARGS'='"--packages" "com.databricks:spark-csv_2.11:1.5.0" "sparkr-shell"')

・ load a csv file as data frame

df <- read.df(sqlContext, inputFilePath, source = "com.databricks.spark.csv", schema = schema)

The full source code is as follows.

・ source code

library(SparkR, lib.loc = c(file.path(Sys.getenv("SPARK_HOME"), "R", "lib")))

Sys.setenv('SPARKR_SUBMIT_ARGS'='"--packages" "com.databricks:spark-csv_2.11:1.5.0" "sparkr-shell"')

# init SparkR
sc <- sparkR.init(appName = "visualization", master = "local[*]")
sqlContext <- sparkRSQL.init(sc)

# parameter
inputFilePath <- "C:/dev/R/test_data.csv"

# schema of data frame
schema <- structType(structField(
                     structField(x = "date", type = "integer", nullable = FALSE),
                     structField(x = "scheme_number", type = "string", nullable = FALSE),
                     structField(x = "product", type = "string", nullable = FALSE),
                     structField(x = "currency", type = "string", nullable = FALSE),
                     structField(x = "sensitivity", type = "double", nullable = FALSE))

# load a csv file as data frame
df <- read.df(sqlContext, inputFilePath, source = "com.databricks.spark.csv", schema = schema)

showDF(df)

2.2. Quiz

Next, I note a simple quiz. Enjoy!! :)

・ quiz

If you don't use the spark-csv package, how do you create the data frame with the folllowing schema from test_data.csv?

・ schema
　 date,scheme_number,product,currency,sensitivity

・ test_data.csv
　 20170228,scheme1,BermudanCallableSwap,JPY,21261339422
　 20170228,scheme2,BermudanCallableSwap,JPY,22759109989
　 20170228,scheme3,BermudanCallableSwap,JPY,21405741891
　 ...

The time limit is 5 minutes :)

3. Answer

The answer is as follows. It's diffuse, but there is a bit of fun :)
You must prepare a record splitter.

・ record splitter

splitRecord <- function(record) {
  Sys.setlocale("LC_ALL", "C")
  part <- strsplit(record, ",")[[1]]
  list(column1 = as.integer(part[1]),
       column2 = part[2],
       column3 = part[3],
       column4 = part[4],
       column5 = as.double(part[5]))
}

And then, you must make a rdd, split it and convert to a data frame.

・ load a csv file and create data frame

# load a csv file
orgData <- SparkR:::textFile(sc, inputFilePath)

# split data
splitData <- SparkR:::lapply(orgData, splitRecord)

# create data frame
df <- SparkR:::createDataFrame(sqlContext, splitData, schema)

The full source code is as follows :)

・ source code

library(SparkR, lib.loc = c(file.path(Sys.getenv("SPARK_HOME"), "R", "lib")))

# init SparkR
sc <- sparkR.init(appName = "visualization", master = "local[*]")
sqlContext <- sparkRSQL.init(sc)

# parameter
inputFilePath <- "C:/dev/R/test_data.csv"

# schema of data frame
schema <- structType(structField(
                     structField(x = "date", type = "integer", nullable = FALSE),
                     structField(x = "scheme_number", type = "string", nullable = FALSE),
                     structField(x = "product", type = "string", nullable = FALSE),
                     structField(x = "currency", type = "string", nullable = FALSE),
                     structField(x = "sensitivity", type = "double", nullable = FALSE))

# record splitter
splitRecord <- function(record) {
  Sys.setlocale("LC_ALL", "C")
  part <- strsplit(record, ",")[[1]]
  list(column1 = as.integer(part[1]),
       column2 = part[2],
       column3 = part[3],
       column4 = part[4],
       column5 = as.double(part[5]))
}

# load a csv file
orgData <- SparkR:::textFile(sc, inputFilePath)

# split data
splitData <- SparkR:::lapply(orgData, splitRecord)

# create data frame
df <- SparkR:::createDataFrame(sqlContext, splitData, schema)

showDF(df)

4. Conclusion

If you enjoy this quiz, I  will be pleased :) The concert will start soon!

・ Gustav Mahler    Symphony No. 5
・ Richard Wagner  Tristan and Isolde: Prelude and Love Death

・ Konzerthausorchester Berlin, cond. Eliahu Inbal
・ Tokyo

SparkR with Visual Studio and RStudio

1. Introduction

I usually see many lovers of R in financial investment banks. If they have the capability of processing  'Big Data' by R that they loves, they may feel so happy :) SparkR makes it possible. In this article, I note how to enjoy SparkR.

・SparkR

　https://github.com/apache/spark/tree/master/R

SparkR is included in Apache Spark. I use RStudio and Visual Studio  as  IDE. The versions of Apache Spark, R, Visual Studio and RStudio are as follows.

Apache Spark 1.6.3
R 3.3.2
RStudio 1.0.136
Visual Studio Community 2015 Update 3 (Update 3 or later are necessary)

　

2. SparkR with IDE

I deployed  Spark 1.6.3 as follows and set SPARK_HOME to C:\enjyoyspace\spark-1.6.3. RStudio is often used in Executing SparkR as IDE. We can also use Visual Studio and there are many lovers of Visual Studio, So, I note how to execute SparkR with RStudio and Visual Studio.

C:\enjyoyspace\spark-1.6.3
    ├─bin
    ├─conf
    ├─data
    ├─ec2
    ├─examples
    ├─lib
    ├─licenses
    ├─python
    ├─R

2.1. Use RStudio

SparkR is loaded as follows. SparkR's APIs are so easy and kind for the lovers of R language. So, learning costs may be very low for them :)

library(SparkR, lib.loc = c(file.path(Sys.getenv("SPARK_HOME"), "R", "lib")))

The full source code is as follows :)

# load SparkR
library(SparkR, lib.loc = c(file.path(Sys.getenv("SPARK_HOME"), "R", "lib")))

sc <- sparkR.init(appName = "CalcProhitAndLoss")
sqlContext <- sparkRSQL.init(sc)

# create data frame (R)
shockedPV <- data.frame(date = c("20161001", "20161101", "20161224"), PV = c(10000000000, 10000001000, 10000002000))
nonShockedPV <- data.frame(date_nonShocked = c("20161001", "20161101", "20161224"), PV_nonShocked = c(9000000000, 9000001000, 9000002000))

# create data frame (SparkR) from data frame (R)
shockedPVforSparkR <- createDataFrame(sqlContext,shockedPV)
nonShockedPVforSparkR <- createDataFrame(sqlContext,nonShockedPV)

# join of RDDs
masterPV <- join(shockedPVforSparkR, nonShockedPVforSparkR, shockedPVforSparkR$date == nonShockedPVforSparkR$date_nonShocked)

# register table
registerTempTable(masterPV, "masterPV")

# SparkSQL
prohitAndLossForSparkR <- sql(sqlContext, "SELECT date, PV-PV_nonShocked AS prohit_and_loss FROM masterPV")

# collect query results
prohitAndLoss <- collect(prohitAndLossForSparkR)

# display collected results
print(prohitAndLoss)

2.2. Use Visual Studio Community

In the execution of SparkR with Visual Studio, it's necessary to install the following plugin.  The version of Visual Studio must be '2015 Update 3 or later'.

・R Tools for Visual Studio
　https://microsoft.github.io/RTVS-docs/

We are able to use the script described in 2.1. There is nothing to change :)

3. Execution Result

The execution result is as follows. The calculation results are noramlly computed :)

      date prohit_and_loss
1 20161224           1e+09
2 20161001           1e+09
3 20161101           1e+09

4. Conclusion

We can use an attractive tool 'SparkR' with Visual Studio or RStudio. If you use Visual Studio, the version of Visual Studio is '2015 Update 3 or later'.

Executing R Script Files from C# with R.NET

1. Introduction

The photo is Suimono :) It includes Ebi Shinjyo, which is a steamed shrimp dumpling.

In this article, I note how to execute R script files from C# and the execution results. I integrated C# and R with R.NET. R.NET has so simple APIs and is easy to use.

・R.NET

　https://github.com/jmp75/rdotnet

The versions of .NET Framework, R.NET and R are as follows.

.NET Framework 4.5
R.NET 1.6.5
R 3.3.2

2. Integrating C# and R

I want to use R to execute machine learning models, display useful charts and so on, and C# to process various parameters and pass them to R. It's because R has the good capabilities of graphical processing and machine learning but isn't good at handling objects.
And I also want to incorporate R script files implemented by other persons such as quantitative analysts and data scientists, into C# applications.

2.1. C# Program

First, I note the C# program. It's required to pass arguments from it to a R script file and run the R script file. These are realized by the following methods. It's so simple :)

・ passing arguments from C# programs to R script files

REngine.SetCommandLineArguments (string[] parameters)

・ running R script files

REngine.Evaluate ("source('script file path')")

The full source code is as follows :)

・ Program.cs

using RDotNet;

namespace FinancialEngineering
{
    class Program
    {
        static void Main(string[] args)
        {
            var scriptFilePath = "C:/dev/R/test_script.R";
            var riskCsvPath = "C:/dev/R/test_data.csv";
            var valueAtRisk = "25750000000";
            ExecuteScriptFile(scriptFilePath, riskCsvPath, valueAtRisk);
        }

        public static void ExecuteScriptFile(string scriptFilePath, string paramForScript1,
                                                                                                     string paramForScript2)
        {
            using (var en = REngine.GetInstance())
            {
                var args_r = new string[2] { paramForScript1, paramForScript2 };
                var execution = "source('" + scriptFilePath + "')";
                en.SetCommandLineArguments(args_r);
                en.Evaluate(execution);
            }
        }
    }
}

'C:/dev/R/test_data.csv' in this program is like the following csv file.

・ test_data.csv

date,risk_value
20160128,21261339422
20160129,22759109989
20160130,21405741891
...

2.2. R Script File

Next, I wrote the following R script file. It's important that the arguments passed from the C# program become available  by 'args <- commandArgs()'. These are stored after 'args[2]'.

・ test_script.R

args <- commandArgs()
riskValueFilepath <- args[2]
VaR <- args[3]

data <- read.csv(riskValueFilepath, header = TRUE)
attach(data)
x <- strptime(data$date, '%Y%m%d', tz = '')

par(xaxt = 'n')

plot(x, data$risk_value, type = 'n', xlab = '',ylab ='')
lines(x, data$risk_value, type = 'l',col = 'thistle4',add = T)
points(x, data$risk_value, type = 'p',pch = 20,col = ifelse(data$risk_value > VaR, 'tomato', 'thistle4'), add = T)

par(xaxt = 's')

y <- as.POSIXct(round(range(x), "days"))
axis.POSIXct(1,at = seq(y[1],y[2],by = "1 day"),format = '%Y%m%d')
title('Test of Integrating C# and R\n(Points above Value at Risk)', xlab = 'Date',ylab = 'Risk Value')

3. Execution Result

The execution result is as follows. It noramlly ran :)

4. Conclusion

R.NET enables you to intagrate C# and R. Because passing arguments from C# to R and running R script files are possible, it's easy to incorporate  R script files into C# applications.

Point to Note in Upgrading Apache Spark's Version to 2.0 and Executing

1. Introduction

To upgrade Spark's version to 2.0 is so nice. It's because the perfprmance of Spark 2.0's Tungsten engine is faster than of 1.6 as the following Databricks's blog 'Faster'.

・Technical Preview of Apache Spark 2.0 Now on Databricks
　Easier, Faster, and Smarter
　https://databricks.com/blog/2016/05/11/apache-spark-2-0-technical-preview-easier-faster-and-smarter.html

In executing a Spark 2.0 application, the spark-submit script which I had used in executing a Spark 1.6 application wasn't available. The purpose of this article is to show it's solution.

2. Problem

My Spark version was 1.6.0. I upgraded Spark's version to 2.0.1 last week. I build my application with Spark 2.0.1 and deploy it. And then, I executed the following spark-submit script.

spark-submit --master yarn-client --class xxxxx --num-executors 8 --executor-memory 16G --executor-cores 20 yyyyy.jar ... 

But the following error was occured.

16/12/01 19:35:21 WARN datasources.DataSource: Error while looking for metadata directory.
org.apache.spark.SparkException: Unable to create database default as failed to create its directory hdfs://namenode:9000/usr/spark-2.0.1-bin-hadoop2.7/spark-warehouse
...

3. Solution

In Spark 2.0, a warehouse directory is used to store tables and 'spark.sql.warehouse.dir' is added to set the warehouse location. If default file system is set to a local file system, avobe error isn't occured.
My default file system was set to HDFS, so hdfs://namenode:9000/usr/spark-2.0.1-bin-hadoop2.7/spark-warehouse was tried to create and failed.

I added the following string to the spark-submit script and explicitly set a local directory path as the default warehouse path.

--conf spark.sql.warehouse.dir=file:///c:\temp\spark-warehouse 

The full of spark-submit script is following. I executed this script and confirmed that 'c:\temp\spark-warehouse' was created and the job successfully terminated.

spark-submit --master yarn-client --conf spark.sql.warehouse.dir=file:///c:\temp\spark-warehouse --class xxxxx --num-executors 8 --executor-memory 16G --executor-cores 20 yyyyy.jar ... 

4. Conclusion

If your default file system is set to HDFS and you execute a Spark 2.0 application, there is a need to set a local directory path as the default warehouse path.

Design of Filling Up Lacking Data with Apache Spark

I considered the design of filling up lacking data with Apache Spark :)

1. Introduction

The category of rows which id is 1, 4, or 5 is A, B and C. Otherwise, the row which id is 2 and category is B doesn't exist, and the rows which Id is 3 and category is A or C don't exist. I needed to fill up these lacking data like following.

The purpose of this article is to puropose a effective design such that the following figure is achieved.

2. The Design

The design which proposed in this article is constructed by two parts.

2.1. Part1

It is inefficient to filter a RDD by id, check that the RDD contains a particular row and fill up lacking data. In order to do efficient check and filling up, I created a RDD of key-value pairs and grouped the elements by a key.


   ・Stage1: create Pair RDD and group by key
      map() -> groupBykey()

2.2. Part2

By groupByKey(), Spark's shuffle work and partitions are created again like the following left figure. If groupByKey() don't work, a computer must search many partitions. After groupByKey(), however, the computer only search one partition whose id is equal to 2 because multiple lows which have same id are divided into the same partition.

In addition, above checking and filling up are pipeline processing, thus the execution time are reduced.


   ・Stage2: check and fill up lacking data
      map() -> flatMap()

3. Conclusion

If you don't want to do full-scan that is to search all partitions (long time), the design proposed in section 2 efficiently works and reduces the execution time greatly.

Architecture for Overnight Batch Processing and Data Scientists' Work with Apache Spark

I'm on vacation and consider architecture for overnight batch processing and data scientists' work with Apache Spark :)

I. Business flow

The following is a simple business flow I assumed. After overnight batch processing, many data scientists start to work :)

II. Architecture with the use of Parquet

They also request the usage of DataFrames created by overnight batch processing and the performance of queries. In this case, I think architecture with the use of Parquet is suitable.

An example of this architecture is shown below.

DataFrame 1 is converted to Parquet 1. When data scientists want to use DataFrame 1, they load it from Parquet 1. DataFrame 4 is also the same. But not all of columns are loaded. Parquet is a columnar storage format so that it allows loading only selective records.

Therefore, the performance of queries against a Parquet table is higher than against a Text table.

III. Usage of Parquet

How to use Parquet is so easy. We can convert a DataFrame to a Parquet file like the following.

・convert a DataFrame to a Parquet file

dataFrame1.write.format("parquet").save("parquetFilePath")

We can also load a DataFrame from a Parquet file like the following.

・load a DataFrame from a Parquet file

val dataFrame1' = sqlContext.read.parquet("parquetFilePath")
                                              // Not all of columns are loaded. Only selective columns are loaded.

val dataFrame5 = dataFrame1'.join(dataFrame3, dataFrame1'("id") === dataFrame3("id"))
.groupBy(dataFrame3("category1"), dataFrame3("category2"))
.agg(avg(dataFrame1'(value)))

Building Apache Spark on Windows 10

I built Apache Spark on Windows 10 and checked the operation of the Spark :)

I. Building

First, I downloaded the Spark source code (v1.6.0.zip) and then extracted all files.

・Apache Spark source code

https://github.com/apache/spark/releases
v1.6.0.zip

・after extracting

C:\enjyoyspace
    ├─spark-1.6.0
    ├─assembly
    ├─bagel
    ├─bin
    ├─build

Next, I compiled with Scala 2.10.6 to produce a Spark package (Hadoop 2.6.0) by using Apache Maven. Apache Maven is a so loveable :) The changes of pom.xml and the command are following.

・changes of pom.xml

C:\enjyoyspace\spark-1.6.0\pom.xml

before:    <scala.version>2.10.5</scala.version>
after:      <scala.version>2.10.6</scala.version>

・command

cd C:\enjyoyspace\spark-1.6.0
mvn -Pyarn -Phadoop-2.6 -Dhadoop.version=2.6.0 -DskipTests clean package

(install Apache Maven: https://maven.apache.org/install.html)

The Spark package (Hadoop 2.6.0) was produced.

・Spark package (Hadoop 2.6.0)

C:\enjyoyspace\spark-1.6.0\assembly\target\scala-2.10\spark-assembly-1.6.0-hadoop2.6.0.jar

II. Operation Check

I checked the operation capabilities of Spark (testing interactive programs). It noramlly ran :)

・command

C:\enjyoyspace\spark-1.6.0>bin\spark-shell

・using Spark

Welcome to
      ____              __
     / __/__  ___ _____/ /__
    _\ \/ _ \/ _ `/ __/  '_/
   /___/ .__/\_,_/_/ /_/\_\   version 1.6.0
      /_/

Using Scala version 2.10.6 (Java HotSpot(TM) 64-Bit Server VM, Java 1.8.0_77)
Type in expressions to have them evaluated.

Type :help for more information.
Spark context available as sc.
SQL context available as sqlContext.

scala> val lines = sc.parallelize(List("Sqoop", "from external datastores into HDFS", "Julia", "Julia is a high-performance dynamic programming language", "JuliaCon 2016"))
lines: org.apache.spark.rdd.RDD[String] = ParallelCollectionRDD[0] at parallelize at< console>:27

scala> lines.count()

res0: Long = 5

scala> val words = lines.flatMap(line => line.split(" "))

words: org.apache.spark.rdd.RDD[String] = MapPartitionsRDD[1] at flatMap at< console>:29

scala> words.count()
res1: Long = 16

scala>

・Spark Web UI (http://localhost:4040/jobs/)