Using Parquet Files

By |

If you’re familiar with Spark, you know that a dataframe is essentially a data structure that contains “tabular” data in memory. That is, it consists of rows and columns of data that can, for example, store the results of an SQL-style query. Dataframes can be saved into HDFS as Parquet files. Parquet files not only preserve the schema information of the dataframe, but will also compress the data when it gets written into HDFS. This means that the saved file will take up less space in HDFS and it will load faster if you read the data again later. Therefore, it is a useful storage format for data you may want to analyze multiple times.

The Pyspark example below uses Reddit data which is available to all Flux Hadoop users in HDFS ‘/var/reddit’. This data consists of information about all posts made on the popular website Reddit, including their score, subreddit, text body, author, all of which can make for interesting data analysis.

#First, launch the pyspark shell

pyspark --master yarn-client --queue <your_queue> --num-executors 35 --executor-cores 4 --executor-memory 5g

#Load the reddit data into a dataframe

>>> reddit = sqlContext.read.json("/var/reddit/RS_2016-0*")

#Set compression type to snappy

>>> sqlContext.setConf("spark.sql.parquet.compression.codec", "snappy")

#Write data into a parquet file - this example puts it into your HDFS home directory as “reddit.parquet”

>>> reddit.write.parquet("reddit.parquet")

#Create a new dataframe from parquet file 

>>> parquetFile = sqlContext.read.parquet("reddit.parquet")

#Register dataframe as a SQL temporary table

>>> parquetFile.registerTempTable(“reddit_table")

#Query the table

#Can really be any query, but this query will find some of the more highly rated posts

>>> ask = sqlContext.sql(“SELECT title FROM reddit_table WHERE score > 1000 and subreddit = ‘AskReddit’”)

#Since we created the dataframe “ask” with the previous query, we can write it out to HDFS as a parquet file so it can be accessed again later

>>> ask.write.parquet(“ask.parquet”)

#Exit the pyspark console - you’ll view the contents of your parquet file after

>>> exit()

 

To view the contents of your Parquet file, use Parquet tools. Parquet tools is a command line tool that aids in the inspection of Parquet files, such as viewing its contents or its schema.

#view the output

hadoop parquet.tools.Main cat ask.parquet

#view the schema; in this case, just the “title” of the askreddit thread

hadoop parquet.tools.Main schema ask.parquet

 

Using mrjob

By | | No Comments

Another way to run Hadoop jobs is through mrjob. Mrjob is useful for testing out smaller data on another system (such as your laptop), and later being able to run it on something larger, like a Hadoop cluster. To run an mrjob on your laptop, you can simply remove the “-r hadoop” from the command in the example we use here.

A classic example is a word count, taken from the official mrjob documentation here.

Save this file as mrjob_test.py.

"""The classic MapReduce job: count the frequency of words.
"""
from mrjob.job import MRJob
import re

WORD_RE = re.compile(r"[\w']+")


class MRWordFreqCount(MRJob):

    def mapper(self, _, line):
        for word in WORD_RE.findall(line):
            yield (word.lower(), 1)

    def combiner(self, word, counts):
        yield (word, sum(counts))

    def reducer(self, word, counts):
        yield (word, sum(counts))


if __name__ == '__main__':
     MRWordFreqCount.run()

Then, run the following command:

python mrjob_test.py -r hadoop /etc/motd

You should receive an output with the word count of the file /etc/motd. You can also try this with any other file you have that contains text.

Using Hive on Spark

By | | No Comments

Running Hive on Spark rather than MapReduce, its default, can be a faster alternative. The easiest way to do this is to set the hive execution engine to spark on your Hive session.

hive --hiveconf mapreduce.job.queuename=<your_queue>
set hive.execution.engine=spark;

Next, you’ll want to set the number of executor instances, executor cores, and executor memory. For a better idea of how to find the ideal settings for this, you can take a look at this documentation. Changing these settings is essential to getting the job to run quickly. An example of doing so would be:

set spark.executor.instances=15;
set spark.executor.cores=4;
set spark.executor.memory=3g;

Then, run a query just like you would for any Hive job (such as the examples earlier in this guide), and it should run (faster) with Spark instead.

For more information on how to use Hive on Spark and its advantages, check out this post.

Using PySpark

By |

Spark and PySpark utilize a container that their developers call a Resilient Distributed Dataset (RDD) for storing and operating on data. The most important characteristic of Spark’s RDD is that it is immutable – once created, the data it contains cannot be updated. New RDDs can be created by transforming the data in another RDD, which is how analysis is done with Spark.

Save this file as job.py.

from pyspark import SparkConf, SparkContext
import sys

# This script takes two arguments, an input and output
if len(sys.argv) != 3:
  print('Usage: ' + sys.argv[0] + ' <in> <out>')
  sys.exit(1)

input = sys.argv[1]
output = sys.argv[2]

# Set up the configuration and job context
conf = SparkConf().setAppName('AnnualWordLength')
sc = SparkContext(conf=conf)


# Read in the dataset and immediately transform all the lines in arrays
data = sc.textFile(input).map(lambda line: line.split('\t'))

# Create the 'length' dataset as mentioned above. This is done using the next two variables, and the 'length' dataset ends up in 'yearlyLength'.
yearlyLengthAll = data.map(
    lambda arr: (int(arr[1]), float(len(arr[0])) * float(arr[2]))
)
yearlyLength = yearlyLengthAll.reduceByKey(lambda a, b: a + b)

# Create the 'words' dataset as mentioned above.
yearlyCount = data.map(
    lambda arr: (int(arr[1]), float(arr[2]))
).reduceByKey(
    lambda a, b: a + b
)

# Create the 'average_length' dataset as mentioned above.
yearlyAvg = yearlyLength.join(yearlyCount).map(
    lambda tup: (tup[0], tup[1][0] / tup[1][1])
)

# Save the results in the specified output directory.
yearlyAvg.saveAsTextFile(output)

# Finally, let Spark know that the job is done.
sc.stop()

This above script averages the lengths of words in the NGrams dataset by year. There are two main operations in the above code: ‘map’ and ‘reduceByKey’. ‘map’ applies a function to each RDD element and returns a new RDD containing the results. ‘reduceByKey’ applies a function to the group of values with the same key – for all keys – and returns an RDD with the result.

The job can be submitted by running:

spark-submit \
 --master yarn-client \
 --queue <your_queue> \
 --num-executors 35 \
 --executor-memory 5g \
 --executor-cores 4 \
 job.py /var/ngrams ngrams-out


hdfs dfs -cat ngrams-out/*

 

The only required arguments from the above job submission command are ‘–master yarn-client’ and ‘–queue <your_queue>’. The values passed to the other arguments may be modified in order to get better performance or conform to the limits of your queue.

*Note: If you want to use Python 3.5 instead of our default 2.7 in your pyspark job, simply run the following commands, and submit your job normally using your Python 3.5 code:

export SPARK_YARN_USER_ENV=PYTHONHASHSEED=0

export PYSPARK_PYTHON=/sw/lsa/centos7/python-anaconda3/created-20170424/bin/python

 

In addition to writing a job and submitting it, Spark comes with an interactive Python console, which can be opened this way:

# Load the pyspark console 
pyspark --master yarn-client --queue <your_queue>

This interactive console can be used for prototyping or debugging.

Using Spark

By |

Spark and PySpark utilize a container called Resilient Distributed Dataset (RDD) for storing and operating on data. The most important characteristic of Spark’s RDD is that it is immutable — once created, the data it contains cannot be updated. New RDDs can be created by transforming the data in another RDD, which is how analysis is done with Spark.

Using Spark’s native language, Scala, requires more setup than using PySpark. Some example Scala jobs, including the same example job in the PySpark documentation, can be found on this website. That Spark code has some trivial set up required to run a Spark job, and all of the actual logic is in the ‘run’ function.

In addition to the example code in the Spark-examples repository is a gradle build file. Gradle is a popular build tool for Java and Scala. This code can be downloaded and built by logging on to flux-login and running:

git clone https://bitbucket.org/umarcts/spark-examples
cd spark-examples
./gradlew jar

The last command, “./gradlew jar”, will download all dependencies, compile the code, run tests, and package all of the code into a Java ARchive (JAR). This JAR is submitted to the cluster to run a job. For example, the AverageNGramLength job can be launched by running:

spark-submit \
   --class com.alectenharmsel.examples.spark.AverageNGramLength \
   --master yarn-client \
   --executor-memory 3g \
   --num-executors 35 \
   --queue <your_queue> \
 build/libs/spark-examples-*-all.jar /var/ngrams ngrams-out

The output will be located in your home directory in a directory called ‘ngrams-out’, and can be viewed by running:

hdfs dfs -cat ngrams-out/* | tail -5

The output should look like this:

spark output

Then, delete your output directory:

hdfs dfs -rm -r -skipTrash /user/<your_uniqname>/ngrams-out

Similar to PySpark’s interactive Python shell, Spark has an interactive Scala shell for prototyping and debugging. The Spark shell can be launched by running:

spark-shell --master yarn-client --queue <your_queue>

Using Streaming (Other Programming Methods)

By |

It is also possible to write a job in any programming language, such as Python or C, that operates on tab-separated key-value pairs. The same example done above with Hive and Pig can also be written in Python and submitted as a Hadoop job using Hadoop Streaming. Submitting a job with Hadoop Streaming requires writing a mapper and a reducer. The mapper reads input line by line and generates key-value pairs for the reducer to “reduce” into some sort of sensible data. For our case, the mapper will read in lines and output the year as the key and a ‘1’ as the value if the ngram in the line it reads has only appeared in a single volume. The python code to do this is:

(Save this file as map.py)

#!/usr/bin/env python2.7
import fileinput
for line in fileinput.input():
 arr = line.split("\t")
 try:
    if int(arr[3]) == 1:
       print("\t".join([arr[1], '1']))
 except IndexError:
       pass
 except ValueError:
       pass

 

Now that the mapper has done this, the reduce merely needs to sum the values based on the key:

(Save this file as red.py)

#!/usr/bin/env python2.7

import fileinput

data = dict()

for line in fileinput.input():
  arr = line.split("\t")
  if arr[0] not in data.keys():
     data[arr[0]] = int(arr[1])
  else:
     data[arr[0]] = data[arr[0]] + int(arr[1])

for key in data:
 print("\t".join([key, str(data[key])]))

 

Submitting this streaming job can be done by running the below command:

yarn jar /usr/lib/hadoop-mapreduce/hadoop-streaming.jar \
 -Dmapreduce.job.queuename=<your_queue> \
 -input /var/ngrams \
 -output ngrams-out \
 -mapper map.py \
 -reducer red.py \
 -file map.py \
 -file red.py \
 -numReduceTasks 10


hdfs dfs -cat ngrams-out/* | tail -5

streaming outputhdfs dfs -rm -r -skipTrash /user/<your_uniqname>/ngrams-out

Using Pig

By |

Pig is similar to Hive and can do the same thing. The Pig code to do this is a little bit longer due to its design. However, writing long Pig code is generally easier that writing multiple SQL queries that that chain together, since Pig’s language, PigLatin, allows for variables and other high-level constructs.

# Open the interactive pig console
pig -Dmapreduce.job.queuename=<your_queue>

# Load the data
ngrams = LOAD '/var/ngrams' USING PigStorage('\t') AS 
(ngram:chararray,
year:int, count:long, volumes:long);

# Look at the schema of the ngrams variable
describe ngrams;

# Count the total number of rows (should be 1201784959)
ngrp = GROUP ngrams ALL;
count = FOREACH ngrp GENERATE COUNT(ngrams);
DUMP count;

# Select the number of words, by year, that have only appeared in a single volume
one_volume = FILTER ngrams BY volumes == 1;
by_year = GROUP one_volume BY year;
yearly_count = FOREACH by_year GENERATE group, COUNT(one_volume);
DUMP yearly_count;

The last few lines of output should look like this:

pig output

More information on Pig can be found on the Apache website.

Using Hive

By |

To demonstrate Hive, below is a short tutorial. The tutorial uses the Google NGrams dataset, which is available in HDFS in /var/ngrams.

# Open the interactive hive console
hive --hiveconf mapreduce.job.queuename=<your_queue>

# Create a table with the Google NGrams data in /var/ngrams
CREATE EXTERNAL TABLE ngrams(ngram STRING, year INT, count 
BIGINT,
     volumes BIGINT)
     ROW FORMAT DELIMITED
     FIELDS TERMINATED BY '\t'
     STORED AS TEXTFILE
     LOCATION '/var/ngrams';

# Look at the schema of the table
DESCRIBE ngrams;

# Count the total number of rows (should be 1201784959)
SELECT COUNT(*) FROM ngrams;

# Select the number of words, by year, that have only appeared in a single volume
SELECT year, COUNT(ngram) FROM ngrams WHERE 
volumes = 1
GROUP BY year;

# Optional: delete your ngrams table
DROP table ngrams;

# Exit the Hive console
QUIT;

The last few lines of output should look something like this:

hive output

More information can be found on the Apache website.

Using Hadoop

By |

Writing Hadoop MapReduce code in Java is the lowest level way to program against a Hadoop cluster. Hadoop’s libraries do not contain any abstractions, like Spark RDDs or a Hive or Pig-like higher level language. All code must implement the MapReduce paradigm.

Some example Java MapReduce code can be found in the ARC-TS Hadoop examples page. There is an example job – AverageNGramLength – to generate a yearly average of all words in Google’s
NGrams dataset that is in HDFS at `/var/ngrams`. This job can be built and launched from flux-hadoop-login by running:

git clone https://bitbucket.org/umarcts/hadoop-examples.git
cd hadoop-examples
./gradlew jar
yarn jar build/libs/hadoop-examples-*-all.jar \
  com.alectenharmsel.examples.hadoop.AverageNGramLength \
  -Dmapreduce.job.queuename=<your_queue> \
  /var/ngrams ngrams-out

# List running jobs
yarn application -list

# Once it is done, view the output
hdfs dfs -cat ngrams-out/* | tail -5

The result of the command will print all of the output to the terminal. This output can be redirected to a file and plotted with any plotting tool, such as R or MATLAB.

The output should look like this:

hadoop output

Please note that it can take around 1-5 minutes for everything to build and launch.

Afterwards, it’s smart to remove the directory your output is in, so that you can use the same name in future examples. To do this, run:

hdfs dfs -rm -r -skipTrash /user/<your_uniqname>/ngrams-out

Overview

By |

Hadoop is designed to scale up from single servers to thousands of machines, each offering local computation and storage. Rather than rely on hardware to deliver high-availability, the library itself is designed to detect and handle failures at the application layer, so delivering a highly-available service on top of a cluster of computers, each of which may be prone to failures. (From hadoop.apache.org)

Flux Hadoop is a technology preview and available at no cost. It may have less technical support than the other Flux services. The Flux Hadoop cluster consists of 12 nodes offering 100TB of HDFS space. The software available is:

Using Flux Hadoop requires a Flux user account but does not require an active Flux allocation. Information on getting a Flux account can be found on the Flux User Guide page. You only need to pay attention to steps 1, 2, and 4 in order to use Flux-Hadoop. Please note that in step 4 rather than logging into flux-login.arc-ts.umich.edu, you need to log into flux-hadoop-login.arc-ts.umich.edu. When applying for a Flux account, mention that you would like to use Hadoop and what queue name you would like. To be placed on an already existing queue, the owner/PI on the queue needs to send an email to hpc-support@umich.eduThis video will help teach you some basic Linux navigation commands if needed.

Hadoop consists of two components; HDFS, a filesystem built for high read speeds, and YARN, a resource manager. HDFS is not a POSIX filesystem, so normal command line tools like “cp” and “mv” will not work. Most of the common tools have been reimplemented for HDFS and can be run using the “hdfs dfs” command. All data must be in HDFS for jobs to be able to read it.

Here are a few basic commands:

# List the contents of your HDFS home directory
hdfs dfs -ls

# Copy local file data.csv to your HDFS home directory
hdfs dfs -put data.csv data.csv

# Copy HDFS file data.csv back to your local home directory
hdfs dfs -get data.csv data2.csv

A complete reference of HDFS commands can be found on the Apache website.

Available software

There are multiple tools available (listed at the top) to perform analysis. Hive transparently transforms SQL queries to jobs that run on the Hadoop cluster, allowing researchers already familiar with SQL to continue using it as their data grows beyond the capacity of a traditional database such as MySQL or PostgreSQL. Pig is similar to Hive, except that its language, PigLatin, is slightly more complicated than SQL. However, PigLatin can be used to express complicated or long transformations and queries that are impossible or difficult to express with SQL.

Hadoop jobs may also be directly written in any programming language. Writing a map-reduce job directly gives more control than using Hive or Pig, but requires knowing or learning a programming language. Working on unstructured data generally requires writing a map-reduce job instead of using Hive or Pig, as those tools can’t work with unstructured data.

Spark and PySpark are completely separate from Hadoop, even though the jobs are run on the Hadoop cluster on files that are stored in HDFS. Spark exposes a single data structure, the Resilient Distributed Dataset (RDD), in the Java, Scala, and Python programming languages. RDDs support multiple operations (map, reduce, filter, etc.), and are held in-memory when possible to increase the speed of the analysis. Spark is a good choice over a native Hadoop job or a Streaming job, as Spark contains a number of convenient functions that Hadoop does not have.

About This User Guide

There are a few things that are important to note about this user guide before you begin:

  • Make sure that you always replace “<your_queue>” with whatever the name of your queue is.
  • Any backslashes (‘\’) in a command indicate that it is actually one long command, but it is broken up simply for the sake of readability.
  • These examples all use the Google Ngrams Dataset. You may want to familiarize yourself with that in order to better understand what kind of data you are running these jobs on. Essentially, its schema is:
    1. word – the word of interest
    2. year – the year of the data
    3. count – the number of times the word has appeared in books this year
    4. volumes – the number of volumes the word has appear in for this year