Hadoop Streaming with Python

Hadoop provides a very rich API interface for developing and running MapReduce jobs in Java, however this is not always everybody’s preference. Hadoop Streaming makes it possible to run MapReduce jobs with any language that can access the standard streams STDIN and STDOUT.

Hadoop Streaming creates the plumbing required to build a full map reduce job out to your cluster so that all you need to do is supply a mapper and reducer that uses STDIN for their input and STDOUT for their output.

In today’s example, we’ll re-implement the word count example with python using streaming.

The mapper

In this case, the mapper’s job is to take a line of text (input) a break it into words. We’ll then write the word along with the number 1 to denote that we’ve counted it.

#!/usr/bin/env python

import sys

def read_input(file):
  '''Splits the lines given to it into words and
     produces a generator'''

  for line in file:
      yield line.split()

def main():
  '''Produces (word,1) pairs for every word 
     encountered on the input'''

  data = read_input(sys.stdin)

  for words in data:
      for word in words:
          print '%s,%d' % (word, 1)

if __name__ == "__main__":
  main()

The reducer

The reducers’ job is to come through and process the output of the map function, perform some aggregative operation over the set and produce an output set on this information. In this example, it’ll take the word and each of the 1’s, accumulating them to form a word count.

#!/usr/bin/env python

from itertools import groupby
from operator import itemgetter
import sys

def parse_output(file):
  '''Parses a single line of output produced 
     by the mapper function'''

  for line in file:
      yield line.rstrip().split(',', 1)

def main():
  data = parse_output(sys.stdin)

  # produce grouped pairs to count
  for current_word, group in groupby(data, itemgetter(0)):
    try:

      # produce the total count      
      total_count = sum(int(count) for current_word, count in group)
    
      # send it out to the output
      print "%s,%d" % (current_word, total_count)
    except ValueError:
      # ignore casting errors
      pass

if __name__ == "__main__":
  main()

input | map | sort | reduce

Before we full scale with this job, we can simulate the work that the Hadoop cluster would do for us by using our shell and pipe indirection to test it out. This is not a scale solution, so make sure you’re only giving it a small set of data. We can really treat this process as:

The Zen and the Art of the Internet should do, just fine.

$ cat zen10.txt | ./mapper.py | sort -k1,1 | ./reducer.py

We can now submit this job to the hadoop cluster like so. Remember, we need access to our source data, mapper and reducer from the namenode where we’ll submit this job from.

Submitting your job on the cluster

First, we need to get our input data in an accessible spot on the cluster.

$ bin/hadoop fs -mkdir /user/hadoop
$ bin/hadoop fs -put /srv/zen10.txt /user/hadoop

Make sure it’s there:

$ bin/hadoop fs -ls /user/hadoop

Found 1 items
-rw-r--r--   1 root supergroup     176012 2015-11-20 23:03 /user/hadoop/zen10.txt

Now, we can run the job.

$ bin/hadoop jar share/hadoop/tools/lib/hadoop-streaming-2.7.0.jar \
             -mapper /src/mapper.py \
             -reducer /src/reducer.py 
             -input /user/hadoop/zen10.txt \
             -output /user/hadoop/zen10-output \

The -mapper and -reducer switches are referring to files on the actual linux node whereas -input and -output are referring to HDFS locations.

Results

The results are now available for you in /user/hadoop/zen10-output.

$ bin/hadoop fs -cat \
             /user/hadoop/zen10-output/part-00000

You should see the results start spraying down the page.

. . .
. . .

vernacular,1
version,10
versions,9
very,13
via,20
vic-20,2
vice,1

. . .
. . .

Limitation

So far, the only limitation that I’ve come across with this method of creating map reduce jobs is that the mapper will only work line-by-line. You can’t treat a single record as information spanning across multiple lines. Having information span across multiple lines in your data file should be a rare use case though.

Creating timer jobs with systemd

Creating and executing timer jobs has traditionally been a task for cron. With the arrival of systemd, this responsibility has been shifted onto services and timers. In today’s post, I’ll walk you through creating a service and timer schedule.

Setup

To accomplish this task, we need two files and a couple of shell commands. The basic method to do this is as follows:

• Create a service definition
• Create a timer definition
• Start and enable the timer

In the example today, I’m going to schedule s3cmd each week to run over a mounted drive to sync with s3.

As we’re working with systemd, everything that we’ll do is a unit file.

Create a service definition

The service definition is a unit file which defines the actual work to be done. The following is placed at /etc/systemd/system/sync-to-s3.service.

[Unit]
Description=Runs the sync script for local file shares to s3

[Service]
Type=oneshot
ExecStart=/usr/bin/sh -c 's3cmd sync --check-md5 --follow-symlinks --verbose /mnt/share/ s3://my-s3-bucket/'

Full particulars on this file structure can be found in the documentation about service unit configuration.

Create a timer definition

The timer definition is also another unit file that defines a schedule. The following is named the same as the above, only it gets a .timer extension at /etc/systemd/system/sync-to-s3.timer.

[Unit]
Description=Schedules the sync of local file shares out to s3

[Timer]
OnCalendar=weekly
OnBootSec=10min

[Install]
WantedBy=multi-user.target

Again the documentation defines the full definition of the timer unit configuration.

The OnCalendar takes a value that needs to be understood by the time span parser, so make sure that it’s valid in accordance with the time span reference.

Start and enable the timer

Now that the service and schedule definitions have been created, we can start up the timer:

sudo systemctl start sync-to-s3.timer
sudo systemctl enable sync-to-s3.timer

Now that you’ve got your job up and running, you get the full feature set that systemd offers, including journald. You can use this to inspect the current or historical run logs from invocations:

sudo journalctl -u sync-to-s3

Using PIG

Pig is a data mining and analysis language that you can use to reason about large data sets. The language works with Hadoop’s MapReduce framework to enable the language to crunch large datasets.

In today’s post, I’ll walk you through:

• Loading input data to HDFS
• Writing and Executing your Pig query
• Exploring output data sets

Get started

First of all, we need our source data. For the purposes of this article, I have a very simple data set consisting of 4 fields all sitting in a CSV format file. The file people.csv then looks like this:

I’ll assume that your source file is sitting on a node in your Hadoop cluster that has access to HDFS. We now create an area for us to store our input data as well as upload our source data to HDFS with the following:

# make a directory under the /user/hadoop folder
# to hold our data, called "demo"
$ bin/hadoop fs -mkdir -p /user/hadoop/demo

# perform the upload into the "demo" folder
$ bin/hadoop fs -put /src/people.csv /user/hadoop/demo

We now confirm that the data is actually sitting there, using the familliar ls command:

$ bin/hadoop fs -ls /user/hadoop/demo

HDFS should respond, showing peoeple.csv. in place

Found 1 items
-rw-r--r--   1 root supergroup         80 2015-11-10 06:27 /user/hadoop/demo/people.csv

Running your queries

Now that our source data has been deployed to HDFS and is available to us, we can fire up Pig. There are two modes that you can run Pig in:

• Local which will operate on local data and not submit map reduce jobs to complete its process
• MapReduce which will use the cluster to perform its work

The local mode is quite handy in testing scenarios where your source data set is small and you’re just looking to test something out quickly. The mapreduce mode is where your information needs to scale to the size of your cluster.

# startup Pig so that it's in mapreduce mode
$ pig -x mapreduce

Now that you’ve got Pig started, you’ll be presented with the grunt> prompt. It’s at this prompt that we can enter in our queries for processing. The following query will load our data set, extract the first (id) column and pump it into an output set.

grunt> A = load '/user/hadoop/demo/people.csv' using PigStorage(',');
grunt> B = foreach A generate $0 as id;
grunt> store B into 'id.out';

The source data set is loaded into A. B then takes all of the values in the first column and writes these to id.out.

Pig will now send your question (or query) off into the compute cluster in the form of a map reduce job. You’ll see the usual fanfare scrolling up the screen from the output of this job submission, and you should be able to follow along on the job control web application for your cluster.

Viewing the result

Once the query has finished its process, you’ll be able to take a look at the result. As this has invoked a map reduce job, you’ll be offered the familiar _SUCCESS file in your output folder to illustrate that your query has run successfully.

$ bin/hadoop fs -ls id.out

You’ll also be given the result in the file part-m-00000.

Found 2 items
-rw-r--r--   1 root supergroup          0 2015-11-10 06:54 id.out/_SUCCESS
-rw-r--r--   1 root supergroup         31 2015-11-10 06:54 id.out/part-m-00000

We can take a look at these results now:

$ bin/hadoop fs -cat id.out/part-m-00000

This is a very simple example of how to run a Pig query on your Hadoop cluster. You can see how these ideas will scale with you as your dataset grows. The example query itself isn’t very complex by any stretch, so now that you know how to execute queries you can read up on Pig latin to tune your query writing craft.

Moving from SQL Server to PostgreSQL

SQL Server and PostgreSQL are both relational database systems and as such share similarities that should allow you to migrate your data structures between them. In today’s post, I’m going to go through the process of migrating SQL Server data schema objects over to PostgreSQL.

Immediate differences

Whilst both relational database systems implement a core set of the standard language, there are implementation-specific features which need special consideration. So long as you are capable of wrangling text in your favorite editor, the conversion task shouldn’t be that hard.

The batch terminator GO gets replaces by a much more familiar ;.

Tables

First thing to do for tables is to generate your create scripts. Make sure that you:

• Turn off DROP statement generation for your objects
• Turn on index and keys generation

To safely qualify the names of objects within the database, SQL Server will surround its object names with square brackets [], so you’ll see definitions like this:

-- generated from SQL Server

CREATE TABLE [dbo].[Table1] (
    [ID]          INT IDENTITY (1, 1) NOT NULL,
    [Name]        VARCHAR (50) NOT NULL
    CONSTRAINT [PK_Table1] PRIMARY KEY CLUSTERED ([ID] ASC)
)

PostgreSQL uses double-quotes on object names and doesn’t use the owner (in the above case [dbo]) to qualify names.

In the above example, Table1 is using IDENTITY on its primary key field ID. This gives us the auto-increment functionality that’s so natural in relational database systems. There is a little extra work in PostgreSQL to emulate this behavior through the use of CREATE SEQUENCE and nextval.

The SQL Server definition above now looks like this for PostgreSQL:

-- migrated for PostgreSQL

CREATE SEQUENCE Table1Serial;

CREATE TABLE Table1 (
    ID          INT NOT NULL DEFAULT nextval('Table1Serial'),
    Name        VARCHAR (50) NOT NULL,
    CONSTRAINT PK_Table1 PRIMARY KEY (ID)
);

Stored Procedures

Stored procedures in SQL Server are considered a much more common citizen in the database world than Stored procedures in PostgreSQL. If your database design hinges on extensive use of stored procedures, you’ll be in for a bit of redevelopment.

Both stored procedures and functions are created using the same syntax in PostgreSQL. The actions that either can perform differ though:

A simple function that will square its input value looks as follows:

CREATE OR REPLACE FUNCTION SquareNum(n INT) RETURNS INT AS $$
  BEGIN
    RETURN n * n;
  END;
  $$ LANGUAGE plpgsql;

This can be invoked using SELECT.

SELECT SquareNum(5);

A more in-depth example involves returning a result set from within a stored procedure. You can do this in an unnamed fashion; you won’t control the name of the cursor coming back.

You can pull out a single record set:

CREATE OR REPLACE FUNCTION retrieve_entries() RETURNS refcursor AS $$
  DECLARE
    ref refcursor;
  BEGIN
    OPEN ref FOR SELECT id, name FROM table1;   
    RETURN ref;                                                       
  END;
  $$ LANGUAGE plpgsql;

Multiple record sets:

CREATE OR REPLACE FUNCTION show_entities_multiple() RETURNS SETOF refcursor AS $$
  DECLARE
    ref1 refcursor;           
    ref2 refcursor;                             
  BEGIN
    OPEN ref1 FOR SELECT id, name FROM table1;   
    RETURN NEXT ref1;                                                                              

    OPEN ref2 FOR SELECT id, name FROM table2;   
    RETURN NEXT ref2;
  END;
  $$ LANGUAGE plpgsql;

Invoking these stored procedures so that you can gather the information being returned, requires you to FETCH these details:

BEGIN;
SELECT retrieve_entities();
FETCH ALL IN "<unnamed portal 2>";
COMMIT;

Re-writing retrieve_entities, we can give the caller the option to name their cursor:

CREATE OR REPLACE FUNCTION retrieve_entities(ref refcursor) RETURNS refcursor AS $$
  BEGIN
    OPEN ref FOR SELECT id, name FROM table1;   
    RETURN ref;
  END;
  $$ LANGUAGE plpgsql;

The invocation of this procedure now requires a name:

BEGIN;
SELECT retrieve_entities('entities_cur');
FETCH ALL IN "entities_cur";
COMMIT;

A much more comprehensive run down of stored procedures/functions can be found here and here.

Views

Views fall into the same category as Tables. The syntax remains very much the same with functions that change between the database platforms.

Detaching running processes in bash

There are quite a few times where I’ve run a command on a remote machine and needed to get out of that machine but leave my command running.

I’ll normally start a job that I know is going to take a while using an ampersand like so:

$ long-running-prog &

Really, the nohup command should also be put on the command line so that the command that you execute will ignore the signal SIGHUP.

$ nohup long-running-prog &
$ exit

If you’re already part-way through a running process, you can get it to continue running in the background (while you make your getaway) by doing the following

$ long-running-prog
CTRL-Z
$ bg
$ disown pid

You use CTRL-Z to suspend the running process. The bg command then gets the program running in the background. You can confirm that it is running in the background with the jobs command. Lastly, using disown detatches the process running in the background from your terminal, so that when you exit your session the process will continue.

The LDP has a great article on job control.