Archive

Archive for the ‘Hadoop’ Category

Beyond MapReduce and Apache Hadoop 2.X with Bikas Saha and Arun Murthy

April 2, 2014 Leave a comment

Episode 20 of the podcast was with Bikas Saha and Arun Murthy.

When I spoke with Arun a year or so a go YARN was NextGen Hadoop and there have been a lot of updates, work done and production experience since.

Besides Yahoo! other multi thousand node clusters have been and are running in production with YARN. These clusters have shown 2x capacity throughput which resulted in reduced cost for hardware (and in some cases being able to shut down co-los) while still gaining performance improvements overall to previous clusters of Hadoop 1.X.

I got to hear about some of what is in 2.4 and coming in 2.5 of Hadoop:

  • Application timeline server repository and api for application specific metrics (Tez, Spark, Whatever).
    • web service API to put and get with some aggregation.
    • plugable nosql store (hbase, accumulo) to scale it.
  • Preemption capacity scheduler.
  • Multiple resource support (CPU, RAM and Disk).
  • Labels tag nodes with labels can be labeled however so some windows and some linux and ask for resources with only those labels with ACLS.
  • Hypervisor support as a key part of the topology.
  • Hoya generalize for YARN (game changer) and now proposed as Slider to the Apache incubator.

We talked about Tez which provides complex DAGs of queries to translate what you want to-do on Hadoop without the work arounds for making it have to run in MapReduce.  MapReduce was not designed to be re-workable out side of the parts of the Job it gave you for Map, Split, Shuffle, Combine, Reduce, Etc and Tez is more expressible exposing a DAG API.

PigHiveQueryOnMR

Now becomes with Tez:

PigHiveQueryOnTez

 

There were also some updates on Hive v13 coming out with sub queries, low latency queries (through Tez), high precision decimal points and more!

Subscribe to the podcast and listen to all of what Bikas and Arun had to say.

/*******************************************
 Joe Stein
 Founder, Principal Consultant
 Big Data Open Source Security LLC
 Twitter: @allthingshadoop
********************************************/

 

 

Impala and SQL on Hadoop

February 22, 2014 Leave a comment

The origins of Impala can be found in F1 – The Fault-Tolerant Distributed RDBMS Supporting Google’s Ad Business.

One of many differences between MapReduce and Impala is in Impala the intermediate data moves from process to process directly instead of storing it on HDFS for processes to get at the data needed for processing.  This provides a HUGE performance advantage and doing so while consuming few cluster resources.   Less hardware to-do more!

impala

There are many advantages to this approach over alternative approaches for querying Hadoop data, including::

  • Thanks to local processing on data nodes, network bottlenecks are avoided.
  • A single, open, and unified metadata store can be utilized.
  • Costly data format conversion is unnecessary and thus no overhead is incurred.
  • All data is immediately query-able, with no delays for ETL.
  • All hardware is utilized for Impala queries as well as for MapReduce.
  • Only a single machine pool is needed to scale.

We encourage you to read the documentation for further exploration!

There are still transformation steps required to optimize the queries but Impala can help to-do this for you with Parquet file format.  Better compression and optimized runtime performance is realized using the ParquetFormat though many other file types are supported.

This and a whole lot more was discussed with Marcel Kornacker the Cloudera Architect behind Impala on Episode 18 of the All Things Hadoop Podcast.

/*******************************************
 Joe Stein
 Founder, Principal Consultant
 Big Data Open Source Security LLC
 Twitter: @allthingshadoop
********************************************/

Using Apache Drill for Large Scale, Interactive, Real-Time Analytic Queries

October 29, 2013 Leave a comment

Episode #17 of the podcast is a talk with Jacques Nadeau  available also on iTunes

Apache Drill http://incubator.apache.org/drill/, a modern interactive query engine that runs on top of Hadoop.

Jacques talked about how Apache Drill is a modern query engine that is meant to be a query layer on top of all big data open source systems. Apache Drill is being designed to make the storage engine as plug-able so it could be the interface for any big data storage engine. The first release came out recently to allow developers to understand the data pipeline.

Leveraging an efficient columnar storage format, an optimistic execution engine and a cache-conscious memory layout, Apache Drill is blazing fast. Coordination, query planning, optimization, scheduling, and execution are all distributed throughout nodes in a system to maximize parallelization.

drill_runtime

Perform interactive analysis on all of your data, including nested and schema-less. Drill supports querying against many different schema-less data sources including HBase, Cassandra and MongoDB. Naturally flat records are included as a special case of nested data.

json

Strongly defined tiers and APIs for straightforward integration with a wide array of technologies.

arch

Subscribe to the podcast and listen to what Jacques had to say.  Available also on iTunes

/*********************************
Joe Stein
Founder, Principal Consultant
Big Data Open Source Security LLC
http://www.stealth.ly
Twitter: @allthingshadoop
**********************************/

Big Data, Open Source and Analytics

August 26, 2013 Leave a comment

Episode #15 of the podcast is a talk with Stefan Groschupf  available also on iTunes

Stefan is the CEO of Datameer and talked about how the company started and where it is now. Founded in 2009 by some of the original contributors to Apache Hadoop, Datameer has grown to a global team, advancing big data analytics. After several implementations of Hadoop analytics solutions at Global 500 companies, the founders were determined to build the next generation analytics application to solve the new use cases created by the explosion of structured and unstructured data. Datameer is the single application for big data analytics by combining data integration, data transformation and data visualization. Customers love us and we work to make Datameer even better each day.

Datameer provides the most complete solution to analyze structured and unstructured data. Not limited by a pre-built schema, the point and click functions means your analytics are only limited by your imagination. Even the most complex nested joins of a large number of datasets can be performed using an interactive dialog. Mix and match analytics and data transformations in unlimited number of data processing pipelines. Leave the raw data untouched.

Datameer turbocharges time series analytics by correlating multiple sets of complex, disparate data. Resulting analytics are endless including correlation of credit card transactions with card holder authorizations, network traffic data, marketing interaction data and many more. The end game is a clear window into the operations of your business, giving you the actionable insights you need to make business decisions.

some alt text

Data is the raw materials of insight and the more data you have, the deeper and broader the possible insights. Not just traditional, transaction data but all types of data so that you can get a complete view of your customers, better understand business processes and improve business performance.

Datameer ignores the limitations of ETL and static schemas to empower business users to integrate data from any source into Hadoop. Pre-built data connector wizards for all common structured and unstructured data sources means that data integration is an easy, three step process of where, what and when.

App Market Infographics

Now you never have to waste precious time by starting from scratch. Anyone can simply browse the Analytics App Market, download an app, connect to data, and get instant results. But why stop there? Every application is completely open so you can customize it, extend it, or even mash it up with other applications to get the insights you need.

Built by data scientists, analysts, or subject matter experts, analytic apps range from horizontal use cases like email and social sentiment analysis to vertical or even product-specific applications like advanced Salesforce.com sales-cycle analysis.

Check out the Datameer app market.

Subscribe to the podcast and listen to what Stefan had to say.  Available also on iTunes

/*********************************
Joe Stein
Founder, Principal Consultant
Big Data Open Source Security LLC
http://www.stealth.ly
Twitter: @allthingshadoop
**********************************/

Categories: Hadoop, Podcast

SQL Compatibility in Hadoop with Hive

August 15, 2013 Leave a comment

Episode #14 of the podcast is a talk with Alan Gates available also on iTunes

The Stinger initiative is a collection of development threads in the Hive community that will deliver 100X performance improvements as well as SQL compatibility.

Fast Interactive Query
An immediate aim of 100x performance increase for Hive is more ambitious than any other effort.
SQL Compatibility
Based on industry standard SQL, the Stinger Initiative improves HiveQL to deliver SQL compatibility.

Apache Hive is the de facto standard for SQL-in-Hadoop today with more enterprises relying on this open source project than any alternative. As Hadoop gains in popularity, enterprise requirements for Hive to become more real time or interactive have evolved… and the Hive community has responded.

He spoke in detail about the Stinger initiative, who is contributing to it, why they decided to improve upon Hive and not create a new system and more.

He talked about how Microsoft is contributing in the open source community to improve upon Hive.

Hadoop is so much more than just SQL, one of the wonderful things about Big Data is the power it brings for users to bring different processing models such as realtime streaming with Storm, Graph processing with Giraph and ETL with Pig and all different things to-do beyond just this SQL compatibility.

Alan also talked about YARN and Tez and the benefits of the Stinger initiative to other Hadoop ecosystem tools too.

Subscribe to the podcast and listen to what Alan had to say.  Available also on iTunes

/*********************************
Joe Stein
Founder, Principal Consultant
Big Data Open Source Security LLC
http://www.stealth.ly
Twitter: @allthingshadoop
**********************************/

Categories: Hadoop, Hive, Podcast

Apache BigTop and how packaging infrastructure binds the Hadoop ecosystem together

August 12, 2013 Leave a comment

Episode #12 of the podcast is a talk with Mark Grover and Roman Shaposhnik  Available also on iTunes

Apache Bigtop is a project for the development of packaging and tests of the Apache Hadoop ecosystem.

The primary goal of Bigtop is to build a community around the packaging and interoperability testing of Hadoop-related projects. This includes testing at various levels (packaging, platform, runtime, upgrade, etc…) developed by a community with a focus on the system as a whole, rather than individual projects.

BigTop makes it easier to deploy Hadoop Ecosystem projects including:

  • Apache Zookeeper

  • Apache Flume

  • Apache HBase

  • Apache Pig

  • Apache Hive

  • Apache Sqoop

  • Apache Oozie

  • Apache Whirr

  • Apache Mahout

  • Apache Solr (SolrCloud)

  • Apache Crunch (incubating)

  • Apache HCatalog

  • Apache Giraph

  • LinkedIn DataFu

  • Cloudera Hue

The list of supported Linux platforms has expanded to include:

  • CentOS/RHEL 5 and 6

  • Fedora 17 and 18

  • SuSE Linux Enterprise 11

  • OpenSUSE 12.2

  • Ubuntu LTS Lucid (10.04) and Precise (12.04)

  • Ubuntu Quantal (12.10)

Subscribe to the podcast and listen to what Mark and Roman had to say.  Available also on iTunes

/*********************************
Joe Stein
Founder, Principal Consultant
Big Data Open Source Security LLC
http://www.stealth.ly
Twitter: @allthingshadoop
**********************************/

Hadoop as a Service cloud platform with the Mortar Framework and Pig

August 9, 2013 Leave a comment

Episode #11 of the podcast is a talk with K Young.  Available also on iTunes

Mortar is the fastest and easiest way to work with Pig and Python on Hadoop in the Cloud.

Mortar’s platform is for everything from joining and cleansing large data sets to machine learning and building recommender systems.

Mortar makes it easy for developers and data scientists to do powerful work with Hadoop. The main advantages of Mortar are:

  • Zero Setup Time: Mortar takes only minutes to set up (or no time at all on the web), and you can start running Pig jobs immediately. No need for painful installation or configuration.
  • Powerful Tooling: Mortar provides a rich suite of tools to aid in Pig development, including the ability to Illustrate a script before running it, and an extremely fast and free local development mode.
  • Elastic Clusters: We spin up Hadoop clusters as you need them, so you don’t have to predict your needs in advance, and you don’t pay for machines you don’t use.
  • Solid Support: Whether the issue is in your script or in Hadoop, we’ll help you figure out a solution.

We talked about the Open Source Mortar Framework and their new Open Source tool for visualizing data while writing Pig scripts called Watchtower

The Mortar Blog has a great video demo on Watchtower.

There are no two ways around it, Hadoop development iterations are slow. Traditional programmers have always had the benefit of re-compiling their app, running it, and seeing the results within seconds. They have near instant validation that what they’re building is actually working. When you’re working with Hadoop, dealing with gigabytes of data, your development iteration time is more like hours.

Subscribe to the podcast and listen to what K Young had to say.  Available also on iTunes

/*********************************
Joe Stein
Founder, Principal Consultant
Big Data Open Source Security LLC
http://www.stealth.ly
Twitter: @allthingshadoop
**********************************/

Categories: Hadoop, Pig, Podcast

Hortonworks HDP1, Apache Hadoop 2.0, NextGen MapReduce (YARN), HDFS Federation and the future of Hadoop with Arun C. Murthy

July 23, 2012 2 comments

Episode #8 of the Podcast is a talk with Arun C. Murthy.

We talked about Hortonworks HDP1, the first release from Hortonworks, Apache Hadoop 2.0, NextGen MapReduce (YARN) and HDFS Federations

subscribe to the podcast and listen to all of what Arun had to share.

Some background to what we discussed:

Hortonworks Data Platform (HDP)

from their website: http://hortonworks.com/products/hortonworksdataplatform/

Hortonworks Data Platform (HDP) is a 100% open source data management platform based on Apache Hadoop. It allows you to load, store, process and manage data in virtually any format and at any scale. As the foundation for the next generation enterprise data architecture, HDP includes all of the necessary components to begin uncovering business insights from the quickly growing streams of data flowing into and throughout your business.

Hortonworks Data Platform is ideal for organizations that want to combine the power and cost-effectiveness of Apache Hadoop with the advanced services required for enterprise deployments. It is also ideal for solution providers that wish to integrate or extend their solutions with an open and extensible Apache Hadoop-based platform.

Key Features
  • Integrated and Tested Package – HDP includes stable versions of all the critical Apache Hadoop components in an integrated and tested package.
  • Easy Installation – HDP includes an installation and provisioning tool with a modern, intuitive user interface.
  • Management and Monitoring Services – HDP includes intuitive dashboards for monitoring your clusters and creating alerts.
  • Data Integration Services – HDP includes Talend Open Studio for Big Data, the leading open source integration tool for easily connecting Hadoop to hundreds of data systems without having to write code.
  • Metadata Services – HDP includes Apache HCatalog, which simplifies data sharing between Hadoop applications and between Hadoop and other data systems.
  • High Availability – HDP has been extended to seamlessly integrate with proven high availability solutions.

Apache Hadoop 2.0

from their website: http://hadoop.apache.org/common/docs/current/

Apache Hadoop 2.x consists of significant improvements over the previous stable release (hadoop-1.x).

Here is a short overview of the improvments to both HDFS and MapReduce.

  • HDFS FederationIn order to scale the name service horizontally, federation uses multiple independent Namenodes/Namespaces. The Namenodes are federated, that is, the Namenodes are independent and don’t require coordination with each other. The datanodes are used as common storage for blocks by all the Namenodes. Each datanode registers with all the Namenodes in the cluster. Datanodes send periodic heartbeats and block reports and handles commands from the Namenodes.More details are available in the HDFS Federation document.
  • MapReduce NextGen aka YARN aka MRv2The new architecture introduced in hadoop-0.23, divides the two major functions of the JobTracker: resource management and job life-cycle management into separate components.The new ResourceManager manages the global assignment of compute resources to applications and the per-application ApplicationMaster manages the application‚Äôs scheduling and coordination.An application is either a single job in the sense of classic MapReduce jobs or a DAG of such jobs.The ResourceManager and per-machine NodeManager daemon, which manages the user processes on that machine, form the computation fabric.The per-application ApplicationMaster is, in effect, a framework specific library and is tasked with negotiating resources from the ResourceManager and working with the NodeManager(s) to execute and monitor the tasks.More details are available in the YARN document.
Getting Started

The Hadoop documentation includes the information you need to get started using Hadoop. Begin with the Single Node Setup which shows you how to set up a single-node Hadoop installation. Then move on to the Cluster Setup to learn how to set up a multi-node Hadoop installation.

Apache Hadoop NextGen MapReduce (YARN)

from their website: http://hadoop.apache.org/common/docs/current/hadoop-yarn/hadoop-yarn-site/YARN.html

MapReduce has undergone a complete overhaul in hadoop-0.23 and we now have, what we call, MapReduce 2.0 (MRv2) or YARN.

The fundamental idea of MRv2 is to split up the two major functionalities of the JobTracker, resource management and job scheduling/monitoring, into separate daemons. The idea is to have a global ResourceManager (RM) and per-application ApplicationMaster (AM). An application is either a single job in the classical sense of Map-Reduce jobs or a DAG of jobs.

The ResourceManager and per-node slave, the NodeManager (NM), form the data-computation framework. The ResourceManager is the ultimate authority that arbitrates resources among all the applications in the system.

The per-application ApplicationMaster is, in effect, a framework specific library and is tasked with negotiating resources from the ResourceManager and working with the NodeManager(s) to execute and monitor the tasks.

MapReduce NextGen Architecture

The ResourceManager has two main components: Scheduler and ApplicationsManager.

The Scheduler is responsible for allocating resources to the various running applications subject to familiar constraints of capacities, queues etc. The Scheduler is pure scheduler in the sense that it performs no monitoring or tracking of status for the application. Also, it offers no guarantees about restarting failed tasks either due to application failure or hardware failures. The Scheduler performs its scheduling function based the resource requirements of the applications; it does so based on the abstract notion of a resource Container which incorporates elements such as memory, cpu, disk, network etc. In the first version, only memory is supported.

The Scheduler has a pluggable policy plug-in, which is responsible for partitioning the cluster resources among the various queues, applications etc. The current Map-Reduce schedulers such as the CapacityScheduler and the FairScheduler would be some examples of the plug-in.

The CapacityScheduler supports hierarchical queues to allow for more predictable sharing of cluster resources

The ApplicationsManager is responsible for accepting job-submissions, negotiating the first container for executing the application specific ApplicationMaster and provides the service for restarting the ApplicationMaster container on failure.

The NodeManager is the per-machine framework agent who is responsible for containers, monitoring their resource usage (cpu, memory, disk, network) and reporting the same to the ResourceManager/Scheduler.

The per-application ApplicationMaster has the responsibility of negotiating appropriate resource containers from the Scheduler, tracking their status and monitoring for progress.

MRV2 maintains API compatibility with previous stable release (hadoop-0.20.205). This means that all Map-Reduce jobs should still run unchanged on top of MRv2 with just a recompile.

HDFS Federation

from their website: http://hadoop.apache.org/common/docs/current/hadoop-yarn/hadoop-yarn-site/Federation.html

Background

HDFS LayersHDFS has two main layers:

  • Namespace
    • Consists of directories, files and blocks
    • It supports all the namespace related file system operations such as create, delete, modify and list files and directories.
  • Block Storage Service has two parts
    • Block Management (which is done in Namenode)
      • Provides datanode cluster membership by handling registrations, and periodic heart beats.
      • Processes block reports and maintains location of blocks.
      • Supports block related operations such as create, delete, modify and get block location.
      • Manages replica placement and replication of a block for under replicated blocks and deletes blocks that are over replicated.
    • Storage – is provided by datanodes by storing blocks on the local file system and allows read/write access.

    The prior HDFS architecture allows only a single namespace for the entire cluster. A single Namenode manages this namespace. HDFS Federation addresses limitation of the prior architecture by adding support multiple Namenodes/namespaces to HDFS file system.

Multiple Namenodes/Namespaces

In order to scale the name service horizontally, federation uses multiple independent Namenodes/namespaces. The Namenodes are federated, that is, the Namenodes are independent and don’t require coordination with each other. The datanodes are used as common storage for blocks by all the Namenodes. Each datanode registers with all the Namenodes in the cluster. Datanodes send periodic heartbeats and block reports and handles commands from the Namenodes.

HDFS Federation ArchitectureBlock Pool

A Block Pool is a set of blocks that belong to a single namespace. Datanodes store blocks for all the block pools in the cluster. It is managed independently of other block pools. This allows a namespace to generate Block IDs for new blocks without the need for coordination with the other namespaces. The failure of a Namenode does not prevent the datanode from serving other Namenodes in the cluster.

A Namespace and its block pool together are called Namespace Volume. It is a self-contained unit of management. When a Namenode/namespace is deleted, the corresponding block pool at the datanodes is deleted. Each namespace volume is upgraded as a unit, during cluster upgrade.

ClusterID

A new identifier ClusterID is added to identify all the nodes in the cluster. When a Namenode is formatted, this identifier is provided or auto generated. This ID should be used for formatting the other Namenodes into the cluster.

Key Benefits

  • Namespace Scalability – HDFS cluster storage scales horizontally but the namespace does not. Large deployments or deployments using lot of small files benefit from scaling the namespace by adding more Namenodes to the cluster
  • Performance – File system operation throughput is limited by a single Namenode in the prior architecture. Adding more Namenodes to the cluster scales the file system read/write operations throughput.
  • Isolation – A single Namenode offers no isolation in multi user environment. An experimental application can overload the Namenode and slow down production critical applications. With multiple Namenodes, different categories of applications and users can be isolated to different namespaces.

subscribe to the podcast and listen to all of what Arun had to share.

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

Hadoop distribution bake-off and my experience with Cloudera and MapR

July 10, 2012 5 comments

A few months back we started to endeavor on a new Hadoop cluster @ medialets

We have been live with Hadoop in production since April 2010 and we are still running CDH2. Our current hosting provider does not have a very ideal implementation for us where our 36 nodes are spread out across an entire data center and 5 networks each with 1 GB link. While there are issues with this type of setup we have been able to organically grow our cluster (started at 4 nodes) which powers 100% of our batch analytics for what is now hundreds of millions of mobile devices.

One of our mapreduce jobs processes 30+ billion objects (about 3 TB of uncompressed data) and takes about 90 minutes to run. This jobs runs all day long contiguously. Each run ingests the data that was received while the previous job was running. One of the primary goals of our new cluster was to reduce the time these type of jobs take without having to make any code changes or increase our investment in hardware. We figured besides the infrastructure changes we needed/wanted to make that running an old version of Hadoop meant that we were not taking advantage of all the awesome work that folks have been putting in over the last 2 years to do things like increasing performance.

So we endeavored to what seems to have been coined as “The Hadoop Distribution Bake-off”. We wanted to not only see how new versions of the Cloudera distribution would be running our jobs but also evaluate other distributions that have emerged since we first started with Hadoop. When we did this Hortonwork’s distribution was not released yet otherwise we would have added them and their distro to the possible choices.

First we found a new vendor to setup a test cluster for us http://www.logicworks.com. It was a four node cluster each with 2GB (1G dual bonded) NIC, 12GB of RAM, 4 x 1TB drives (using 3 of the drives for data and one for the OS) and 2x Westmere 5645 2.4GHz Hex-Core CPU. While this was not going to be the exact configuration we were going to end up with it was what they had in inventory and for the purposes of this test it was all about keeping the same hardware running with the same job with the same data and only changing the distro and configurations. As part of our due diligence, performance was not the only point we were interested in but was the primary goal of the bake-off and testing. We also reviewed other aspects of the distributions and companies which ultimately led to our final decision to go with CDH4 for our new cluster.

First, we wanted to create a baseline to see how our data and job did with the existing distribution (CDH2) we run in production with our existing production configuration. Next we wanted to give MapR a shot. We engaged with their team and they spent their time and assistance to help configure and optimize for the job’s test run. Once that was done we wanted to give CDH3 and CDH4 (which was still beta at the time) and the Cloudera folks also lent their time and helped configure and optimize the cluster.

CDH2 = 12 hours 12 min (our production configuration)
MapR = 4 hours 31 min (configuration done by MapR team)
CDH3 = 6 hours 8 min (our production configuration)
CDH4 = 4 hours 20 min (configuration done by Cloudera team)

This told us that the decision between running CDH4 or MapR was not going to be made based on performance of the distribution with our data and mapreduce jobs.

So, we had to look at the other things that were important to us.

MapR has a couple of a really nice features that are unique to their platform. Their file system features with NFS and Snapshots, both are cool so lets go through them quickly. MapR’s underlying proprietary file system allows for these unique features in the Hadoop ecosystem. The NFS feature basically allows you to copy to an NFS share that is distributed across the entire cluster (with a VIP so highly available). This means that you can use the cluster for saving data from your applications and then without any additional copies map-reduce over it. Data is compressible under the hood though this did not mean much to us since we compress all of our data in sequence files using compress by block size on the sequence file. Snapshots (and mirroring to other clusters of those snapshots) is nifty. Being able to take a point in time instance cut of things makes the cluster feel and operate like our SAN. While snapshots are nifty the same end result is capable with a distcp which sure takes longer but is still technically feasible not a lot of other benefits for us or our business, nifty none the less. The main issue we had with all of this was that all of the features that were attractive required us to license their product. Their product also is not open source so we would not be able to build the code, make changes or anything else always having to rely on them for support and maintenance. We met a lot of great folks from MapR but only 2 of them were Apache committers (they may have more on staff, I only met two though) and this is important to us from a support & maintenance perspective… for them it probably is not a huge deal since their platform is not open source and proprietary ( I think I just repeated myself here but did so on purpose ).

Cloudera… tried, true and trusted (I have been running CDH2 for 2 years in production without ever having to upgrade) and know lots of folks that can say the same thing. Everything is Open Source with a very healthy and active community. A handful of times this has been very helpful in development cycles for me to see what the container I was running in was doing to help me resolve the problems I was finding in my own code… or even to simply shoot a question over the mailing list to get a response to a question. As far as the distribution goes, it costs nothing to get it running and have it run in production with all of the features we wanted. If we ever decided to pay for support there are a boat load (a large boat) of Apache Committers not just to the Hadoop project but to lots of projects within the Hadoop eco system all of which are available and part and parcel to help answer questions and make code changes, etc. The philosophy of their distribution (besides just being open source) is to cherry pick changes from Apache Hadoop as soon as they can (or should or want) to be introduced to making their distribution best.

I can think of a lot of industries and companies were MapR would be a good choice over Cloudera.

We decided what was best for us was to go with CDH4 for our new cluster. And, if we ever decide to purchase support we would get it from Cloudera.

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

Categories: Hadoop

Unified analytics and large scale machine learning with Milind Bhandarkar

June 1, 2012 1 comment

Episode #7 of the Podcast is a talk with Milind Bhandarkar.

We talked about unified analytics, machine learning, data science, some great history of Hadoop, the future of Hadoop and a lot more!

subscribe to the podcast and listen to all of what Milind had to share.

/*
Joe Stein
http://www.medialets.com
*/

Hadoop Streaming Made Simple using Joins and Keys with Python

December 16, 2011 10 comments

There are a lot of different ways to write MapReduce jobs!!!

Sample code for this post https://github.com/joestein/amaunet

I find streaming scripts a good way to interrogate data sets (especially when I have not worked with them yet or are creating new ones) and enjoy the lifecycle when the initial elaboration of the data sets lead to the construction of the finalized scripts for an entire job (or series of jobs as is often the case).

When doing streaming with Hadoop you do have a few library options.  If you are a Ruby programmer then wukong is awesome! For Python programmers you can use dumbo and more recently released mrjob.

I like working under the hood myself and getting down and dirty with the data and here is how you can too.

Lets start first with defining two simple sample data sets.

Data set 1:  countries.dat

name|key

United States|US
Canada|CA
United Kingdom|UK
Italy|IT

Data set 2: customers.dat

name|type|country

Alice Bob|not bad|US
Sam Sneed|valued|CA
Jon Sneed|valued|CA
Arnold Wesise|not so good|UK
Henry Bob|not bad|US
Yo Yo Ma|not so good|CA
Jon York|valued|CA
Alex Ball|valued|UK
Jim Davis|not so bad|JA

The requirements: you need to find out grouped by type of customer how many of each type are in each country with the name of the country listed in the countries.dat in the final result (and not the 2 digit country name).

To-do this you need to:

1) Join the data sets
2) Key on country
3) Count type of customer per country
4) Output the results

So first lets code up a quick mapper called smplMapper.py (you can decide if smpl is short for simple or sample).

Now in coding the mapper and reducer in Python the basics are explained nicely here http://www.michael-noll.com/tutorials/writing-an-hadoop-mapreduce-program-in-python/ but I am going to dive a bit deeper to tackle our example with some more tactics.

#!/usr/bin/env python

import sys

# input comes from STDIN (standard input)
for line in sys.stdin:
	try: #sometimes bad data can cause errors use this how you like to deal with lint and bad data
        
		personName = "-1" #default sorted as first
		personType = "-1" #default sorted as first
		countryName = "-1" #default sorted as first
		country2digit = "-1" #default sorted as first
		
		# remove leading and trailing whitespace
		line = line.strip()
	 	
		splits = line.split("|")
		
		if len(splits) == 2: #country data
			countryName = splits[0]
			country2digit = splits[1]
		else: #people data
			personName = splits[0]
			personType = splits[1]
			country2digit = splits[2]			
		
		print '%s^%s^%s^%s' % (country2digit,personType,personName,countryName)
	except: #errors are going to make your job fail which you may or may not want
		pass

Don’t forget:

chmod a+x smplMapper.py

Great! We just took care of #1 but time to test and see what is going to the reducer.

From the command line run:

cat customers.dat countries.dat|./smplMapper.py|sort

Which will result in:

CA^-1^-1^Canada
CA^not so good^Yo Yo Ma^-1
CA^valued^Jon Sneed^-1
CA^valued^Jon York^-1
CA^valued^Sam Sneed^-1
IT^-1^-1^Italy
JA^not so bad^Jim Davis^-1
UK^-1^-1^United Kingdom
UK^not so good^Arnold Wesise^-1
UK^valued^Alex Ball^-1
US^-1^-1^United States
US^not bad^Alice Bob^-1
US^not bad^Henry Bob^-1

Notice how this is sorted so the country is first and the people in that country after it (so we can grab the correct country name as we loop) and with the type of customer also sorted (but within country) so we can properly count the types within the country. =8^)

Let us hold off on #2 for a moment (just hang in there it will all come together soon I promise) and get smplReducer.py working first.

#!/usr/bin/env python
 
import sys
 
# maps words to their counts
foundKey = ""
foundValue = ""
isFirst = 1
currentCount = 0
currentCountry2digit = "-1"
currentCountryName = "-1"
isCountryMappingLine = False

# input comes from STDIN
for line in sys.stdin:
	# remove leading and trailing whitespace
	line = line.strip()
	
	try:
		# parse the input we got from mapper.py
		country2digit,personType,personName,countryName = line.split('^')
		
		#the first line should be a mapping line, otherwise we need to set the currentCountryName to not known
		if personName == "-1": #this is a new country which may or may not have people in it
			currentCountryName = countryName
			currentCountry2digit = country2digit
			isCountryMappingLine = True
		else:
			isCountryMappingLine = False # this is a person we want to count
		
		if not isCountryMappingLine: #we only want to count people but use the country line to get the right name 

			#first check to see if the 2digit country info matches up, might be unkown country
			if currentCountry2digit != country2digit:
				currentCountry2digit = country2digit
				currentCountryName = '%s - Unkown Country' % currentCountry2digit
			
			currentKey = '%s\t%s' % (currentCountryName,personType) 
			
			if foundKey != currentKey: #new combo of keys to count
				if isFirst == 0:
					print '%s\t%s' % (foundKey,currentCount)
					currentCount = 0 #reset the count
				else:
					isFirst = 0
			
				foundKey = currentKey #make the found key what we see so when we loop again can see if we increment or print out
			
			currentCount += 1 # we increment anything not in the map list
	except:
		pass

try:
	print '%s\t%s' % (foundKey,currentCount)
except:
	pass

Don’t forget:

chmod a+x smplReducer.py

And then run:

cat customers.dat countries.dat|./smplMapper.py|sort|./smplReducer.py

And voila!

Canada	not so good	1
Canada	valued	3
JA - Unkown Country	not so bad	1
United Kingdom	not so good	1
United Kingdom	valued	1
United States	not bad	2

So now #3 and #4 are done but what about #2? 

First put the files into Hadoop:

hadoop fs -put ~/mayo/customers.dat .
hadoop fs -put ~/mayo/countries.dat .

And now run it like this (assuming you are running as hadoop in the bin directory):

hadoop jar ../contrib/streaming/hadoop-0.20.1+169.89-streaming.jar -D mapred.reduce.tasks=4 -file ~/mayo/smplMapper.py -mapper smplMapper.py -file ~/mayo/smplReducer.py -reducer smplReducer.py -input customers.dat -input countries.dat -output mayo -partitioner org.apache.hadoop.mapred.lib.KeyFieldBasedPartitioner -jobconf stream.map.output.field.separator=^ -jobconf stream.num.map.output.key.fields=4 -jobconf map.output.key.field.separator=^ -jobconf num.key.fields.for.partition=1

Let us look at what we did:

hadoop fs -cat mayo/part*

Which results in:

Canada	not so good	1
Canada	valued	3
United Kingdom	not so good	1
United Kingdom	valued	1
United States	not bad	2
JA - Unkown Country	not so bad	1

So #2 is the partioner KeyFieldBasedPartitioner explained here further Hadoop Wiki On Streaming which allows the key to be whatever set of columns you output (in our case by country) configurable by the command line options and the rest of the values are sorted within that key and sent to the reducer together by key.

And there you go … Simple Python Scripting Implementing Streaming in Hadoop.  

Grab the tar here and give it a spin.

/*
Joe Stein
Twitter: @allthingshadoop
Connect: On Linked In
*/

Categories: Hadoop, MapReduce, Python

Faster Datanodes with less wait io using df instead of du

May 20, 2011 1 comment

I have noticed often that the check Hadoop uses to calculate usage for the data nodes causes a fair amount of wait io on them driving up load.

Every cycle we can get from every spindle we want!

So I came up with a nice little hack to use df instead of du.

Here is basically what I did so you can do it too.


mv /usr/bin/du /usr/bin/bak_du
vi /usr/bin/du

and save this inside of it

#!/bin/sh

mydf=$(df $2 | grep -vE '^Filesystem|tmpfs|cdrom' | awk '{ print $3 }')
echo -e "$mydf\t$2"

then give it execute permission

chmod a+x /usr/bin/du

restart you data node check the log for no errors and make sure it starts back up

viola

Now when Hadoop calls “du -sk /yourhdfslocation” it will be expedient with its results

whats wrong with this?

1) I assume you have nothing else on your disks that you are storing so df is really close to du since almost all of your data is in HDFS

2) If you have more than 1 volume holding your hdfs blocks this is not exactly accurate so you are skewing the size of each vol by only calculating one of them and using that result for the others…. this is simple to fix just parse your df result differently and use the path passed into the second paramater to know which vol to grep in your df result… your first volume is going to be larger anyways most likely and you should be monitoring disk space another way so it is not going to be very harmefull if you just check and report the first volume’s size

3) you might not have your HDFS blocks on your first volume …. see #2 you can just grep the volume you want to report

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

Categories: Hadoop

Cloudera, Yahoo and the Apache Hadoop Community Security Branch Release Update

May 5, 2011 1 comment

In the wake of Yahoo! having announced that they would discontinue their Hadoop distribution and focus their efforts into Apache Hadoop http://yhoo.it/i9Ww8W the landscape has become tumultuous.

Yahoo! engineers have spent their time and effort contributing back to the Apache Hadoop security branch (branch of 0.20) and have proposed release candidates.

Currently being voted and discussed is “Release candidate 0.20.203.0-rc1″. If you are following the VOTE and the DISCUSSION then maybe you are like me it just cannot be done without a bowl of popcorn before opening the emails. It is getting heated in a good and constructive kind of way. http://mail-archives.apache.org/mod_mbox/hadoop-general/201105.mbox/thread there are already more emails in 5 days of May than there were in all of April. woot!

My take? Has it become Cloudera vs Yahoo! and Apache Hadoop releases will become fragmented because of it? Well, it is kind of like that already. 0.21 is the latest and can anyone that is not a committer quickly know or find out the difference between that and the other release branches? It is esoteric :( 0.22 is right around the corner too which is a release from trunk.

Lets take HBase as an example (a Hadoop project). Do you know what version of HDFS releases can support HBase in production without losing data? If you do then maybe you don’t realize that many people still don’t even know about the branch. And, now that CDH3 is out you can use that (thanks Cloudera!) otherwise it is highly recommended to not be in production with HBase unless you use the append branch http://svn.apache.org/viewvc/hadoop/common/branches/branch-0.20-append/ of 0.20 which makes you miss out on other changes in trunk releases…

__ eyes crossing inwards and sideways with what branch does what and when the trunk release has everything __

Hadoop is becoming an a la cart which features and fixes can I live without for all of what I really need to deploy … or requiring companies to hire a committer … or a bunch of folks that do nothing but Hadoop day in and day out (sounds like Oracle, ahhhhhh)… or going with the Cloudera Distribution (which is what I do and don’t look back). The barrier to entry feels like it has increased over the last year. However, stepping back from that the system overall has had a lot of improvements! A lot of great work by a lot of dedicated folks putting in their time and effort towards making Hadoop (in whatever form the elephant stampedes through its data) a reality.

Big shops that have teams of “Hadoop Engineers” (Yahoo, Facebook, eBay, LinkedIn, etc) with contributors and/or committers on that team should not have lots of impact because ultimately they are able to role their own releases for whatever they need/want themselves in production and just support it. Not all are so endowed.

Now, all of that having been said I write this because the discussion is REALLY good and has a lot of folks (including those from Yahoo! and Cloudera) bringing up pain points and proposing some great solutions that hopefully will contribute to the continued growth and success of the Apache Hadoop Community http://hadoop.apache.org/…. still if you want to run it in your company (and don’t have a committer on staff) then go download CDH3 http://www.cloudera.com it will get you going with the latest and greatest of all the releases, branches, etc, etc, etc. Great documentation too!

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

NoSQL HBase and Hadoop with Todd Lipcon from Cloudera

September 6, 2010 3 comments

Episode #6 of the Podcast is a talk with Todd Lipcon from Cloudera discussing HBase.

We talked about NoSQL and how it should stand for “Not Only SQL” and the tight integration between Hadoop and HBase and how systems like Cassandra (which is eventually consistent and not strongly consistent like HBase) is complementary as these systems have applicability within big data eco system depending on your use cases.

With the strong consistency of HBase you get features like incrementing counters and the tight integration with Hadoop means faster loads with HDFS thanks to a new feature in the 0.89 development preview release in the doc folders called “bulk loads”.

We covered a lot more unique features, talked about more of what is coming in upcoming releases as well as some tips with HBase so subscribe to the podcast and listen to all of what Todd had to say.

/*
Joe Stein
http://www.medialets.com
*/

Hadoop Development Tools By Karmasphere

June 29, 2010 1 comment

In Episode #5 of the Hadoop Podcast http://allthingshadoop.com/podcast/ I speak with Shevek, the CTO of Karmasphere http://karmasphere.com/.  To subscribe to the Podcast click here.

We talk a bit about their existing Community Edition (support Netbeans & Eclipse)

  • For developing, debugging and deploying Hadoop Jobs
  • Desktop MapReduce Prototyping
  • GUI to manipulate clusters, file systems and jobs
  • Easy deployment to any Hadoop version, any distribution in any cloud
  • Works through firewalls

As well as the new products they have launched:

Karmasphere Client:

The Karmasphere Client is a cross platform library for ensuring MapReduce jobs can work from any desktop environment to any Hadoop cluster in any enterprise data network. By isolating the Big Data professional and version of Hadoop, Karmasphere Client simplifies the process of switching between data centers and the cloud and enables Hadoop jobs to be independent of the version of the underlying cluster.

Unlike the standard Hadoop client , Karmasphere Client works from Microsoft Windows as well as Linux and MacOS, and works through SSH-based firewalls. Karmasphere Client provides a cloud-independent environment that makes it easy and predictable to maintain a business operation reliant on Hadoop.

  • Ensures Hadoop distribution and version independence
  • Works from Windows (unlike Hadoop Client)
  • Supports any cloud environment: public, private or public cloud service.
  • Provides:
    • Job portability
    • Operating system portability
    • Firewall hopping
    • Fault tolerant API
    • Synchronous and Asynchronous API
    • Clean Object Oriented Design
  • Making it easy and predictable to maintain a business operation reliant on Hadoop

Karmasphere Studio Professional Edition

Karmasphere Studio Professional Edition includes all the functionality of the Community Edition, plus a range of deeper functionality required to simplify the developer’s task of making a MapReduce job robust, efficient and production-ready.

For a MapReduce job to be robust, its functioning on the cluster has to be well understood in terms of time, processing, and storage requirements, as well as in terms of its behavior when implemented within well-defined “bounds.” Karmasphere Studio Professional Edition incorporates the tools and a predefined set of rules that make it easy for the developer to understand how his or her job is performing on the cluster and where there is room for improvement.

  • Enhanced cluster visualization and debugging
    • Execution diagnostics
    • Job performance timelines
    • Job charting
    • Job profiling
  • Job Export
    • For easy production deployment
  • Support

Karmasphere Studio Analyst Edition

  • SQL interface for ad hoc analysis
  • Karmasphere Application Framework + Hive + GUI =
    • No cluster changes
    • Works over proxies and firewalls
    • Integrated Hadoop monitoring Interactive syntax checking
    • Detailed diagnostics
    • Enhanced schema browser
    • Full JDBC4 compliance
    • Multi-threaded & concurrent

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

Categories: Hadoop, Podcast, Tools

Hadoop and Pig with Alan Gates from Yahoo

Episode 4 of our Podcast is with Alan Gates, Senior Software Engineer @ Yahoo! and Pig committer. Click here to listen.

Hadoop is a really important part of Yahoo’s infrastructure because processing and analyzing big data is increasingly important for their business. Hadoop allows Yahoo to connect their consumer products with their advertisers and users for a better user experience. They have been involved with Hadoop for many years now and have their own distribution. Yahoo also sponsors/hosts a user group meeting which has grown to hundreds of attendees every month.

We talked about what Pig is now, the future of Pig and other projects like Oozie http://github.com/tucu00/oozie1 which Yahoo uses (and is open source) for workflow of MapReduce & Pig script automation. We also talked about Zebra http://wiki.apache.org/pig/zebra, Owl http://wiki.apache.org/pig/owl, and Elephant Bird http://github.com/kevinweil/elephant-bird

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

Ruby Streaming for Hadoop with Wukong a talk with Flip Kromer from Infochimps

Another great discussion on our PodcastClick here to listen.  For this episode our guest was Flip Kromer from Infochimps http://www.infochimps.org.  Infochimps.org’s mission is to increase the world’s access to structured data.  They have been working since the start of 2008 to build the world’s most interesting data commons, and since the start of 2009 to build the world’s first data marketplace. Our founding team consists of two physicists (Flip Kromer and Dhruv Bansal) and one entrepreneur (Joseph Kelly).

We talked about Ruby streaming with Hadoop and why to use the open source project Wukong to simplify implementation of Hadoop using Ruby.  There are some great examples http://github.com/infochimps/wukong/tree/master/examples that are just awesome like the web log analysis that creates the paths (chain of pages) that users go through during their visited session.

It was interesting to learn some of the new implementations and projects that he has going on like using Cassandra to help with storing unique values for social network analysis.  This new project is called Cluster Chef http://github.com/infochimps/cluster_chef.  ClusterChef will help you create a scalable, efficient compute cluster in the cloud. It has recipes for Hadoop, Cassandra, NFS and more — use as many or as few as you like.

  • A small 1-5 node cluster for development or just to play around with Hadoop or Cassandra
  • A spot-priced, ebs-backed cluster for unattended computing at rock-bottom prices
  • A large 30+ machine cluster with multiple EBS volumes per node running Hadoop and Cassandra, with optional NFS for
  • With Chef, you declare a final state for each node, not a procedure to follow. Adminstration is more efficient, robust and maintainable.
  • You get a nice central dashboard to manage clients
  • You can easily roll out configuration changes across all your machines
  • Chef is actively developed and has well-written recipes for webservers, databases, development tools, and a ton of different software packages.
  • Poolparty makes creating amazon cloud machines concise and easy: you can specify spot instances, ebs-backed volumes, disable-api-termination, and more.
  • Hadoop
  • NFS
  • Persistent HDFS on EBS volumes
  • Zookeeper (in progress)
  • Cassandra (in progress)

Another couple of good links we got from Flip were Peter Norvig’s “Unreasonable Effectiveness of Data” thing I mentioned: http://bit.ly/effectofdatabit.ly/norvigtalk

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

Hadoop, BigData and Cassandra with Jonathan Ellis

Today I spoke with Jonathan Ellis who is the Project Chair of the Apache Cassandra project http://cassandra.apache.org/ and co-founder of Riptano, the source for professional Cassandra support http://riptano.com.  It was a great discussion about Hadoop, BigData, Cassandra and Open Source.

We talked about the recent Cassandra 0.6 NoSQL integration and support for Hadoop Map/Reduce against the data stored in Cassandra and some of what is coming up in the 0.7 release.

We touched on how Pig is currently supported and why the motivation for Hive integration may not have any support with Cassandra in the future.

We also got a bit into a discussion of HBase vs Cassandra and some of the benefits & drawbacks as they live in your ecosystem (e.g. HBase is to OLAP as Cassandra is to OLTP).

This was the second Podcast and you can click here to listen.

/*
Joe Stein
http://www.linkedin.com/in/charmalloc/
*/

Making Hadoop and MapReduce easier with Karmasphere

For those folks either just getting started or even already in the the daily trenches of M/R development every day Karmasphere has come about to help developers and technical professionals make Hadoop MapReduce easier http://www.karmasphere.com/. Karmasphere Studio is a desktop IDE for graphically prototyping MapReduce jobs and deploying, monitoring and debugging them on Hadoop clusters in private and public clouds.

* Runs on Linux, Apple Mac OS and Windows.
* Works with all major distributions and versions of Hadoop including Apache, Yahoo! and Cloudera.
* Works with Amazon Elastic MapReduce.
* Supports local, networked, HDFS and Amazon S3 file systems.
* Support for Cascading
* Enables job submission from all major platforms including Windows.
* Operates with clusters and file systems behind firewalls.

So, what can you do with it?

  • Prototype on the desktop: Get going with MapReduce job development quickly. No need for a cluster since Hadoop emulation is included.
  • Deploy to a private or cloud-based cluster: Whether you’re using a cluster in your own network or a cloud, deploy your job/s easily.
  • Debug on the cluster: One of the most challenging areas in MapReduce programming is debugging your job on the cluster. Visual tools deliver real-time insight into your job, including support for viewing and charting Hadoop job and task counters.
  • Graphically visualize and manipulate: Whether it’s clusters, file systems, job configuration, counters, log files or other debugging information, save time and get better insight by accessing it all in one place.
  • Monitor and analyze your jobs in real-time: Get realtime, workflow view of inputs, outputs and intermediate results including map, partition, sort and reduce phases.

Whether you’re new to Hadoop and want to easily explore MapReduce programming or you like the sound of something that helps you prototype, deploy and manage in an integrated environment or you’re already using Hadoop but could use a lot more insight into your jobs running on a cluster, there’s something here for you.

All you need is NetBeans (version 6.7 or 6.8) and Java 1.6 and you’ll be ready to give Karmasphere Studio a whirl.

You do NOT need any kind of Hadoop cluster set up to begin prototyping. But when you are ready to deploy your job on a large data set, you’ll need a virtual or real cluster in your data center or a public cloud such as Amazon Web Services.

An Eclipse version is in progress.

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

Categories: Hadoop, MapReduce

Tips, Tricks And Pointers When Setting Up Your First Hadoop Cluster To Run Map Reduce Jobs

April 28, 2010 7 comments

So you are ready now for your first real cluster?  Have you gone through all the stand alone and pseudo setup of Hadoop?  If not you could/should get familiar http://hadoop.apache.org/common/docs/current/quickstart.html.  The Cloudera distribution has a nice walk through http://archive.cloudera.com/docs/cdh.html.  Cloudera also has great training videos online http://www.cloudera.com/resources/?type=Training.  Also, if you have not already read the Hadoop bible you should (that would be Tom White’s Definitive Guide to Hadoop).

So now that you are setting up your cluster and looking to run some MR (Map Reduce) jobs on it there are a few points of interest to note that deviate from default settings and nuances.  Granted you would eventually end up learning these over a little bit of pain, trial & error, etc but I figure to share the knowledge for what is pretty much run of the mill stuff.

The first episode of our Podcast is also on this topic.

SSH Authentication

In a cluster, the namenode needs to communicate with all the other nodes.  I have already blogged on this previously http://wp.me/pTu1i-29 but it is important to getting you going so I wanted to mention it here again.

Namenode Configuration Changes

The image from your name node is the meta data for knowing where all of the blocks for all of your files in HDFS across your entire cluster are stored.  If you lose the image file your data is NOT recoverable (as the nodes will have no idea what block is for what file).  No fear you can just save a copy to an NFS share.  After you have mounted it then update your hdfs-site.xml configuration.  The property dfs.name.dir takes a comma separated list of directories to save your image to, add it.

Max tasks can be setup on the Namenode or overridden (and set final) on data nodes that may have different hardware configurations.  The max tasks are setup for both mappers and reducers.  To calculate this it is based on CPU (cores) and the amount of RAM you have and also the JVM max you setup in mapred.child.java.opts (the default is 200).  The Datanode and Tasktracker each are set to 1GB so for a 8GB machine the mapred.tasktracker.map.tasks.maximum could be set to 7  and the mapred.tasktracker.reduce.tasks.maximum set to 7 with the mapred.child.java.opts set to -400Xmx (assuming 8 cores).  Please note these task maxes are as much done by your CPU if you only have 1 CPU with 1 core then it is time to get new hardware for your data node or set the mask tasks to 1.  If you have 1 CPU with 4 cores then setting map to 3 and reduce to 3 would be good (saving 1 core for the daemon).

By default there is only one reducer and you need to configure mapred.reduce.tasks to be more than one.  This value should be somewhere between .95 and 1.75 times the number of maximum tasks per node times the number of data nodes.  So if you have 3 data nodes and it is setup max tasks of 7 then configure this between 25 and 36.

As of version 0.19 Hadoop is able to reuse the JVM for multiple tasks.  This is kind of important to not waste time starting up and tearing down the JVM itself and is configurable here mapred.job.reuse.jvm.num.tasks.

Another good configuration addition is mapred.compress.map.output.  Setting this value to true should (balance between the time to compress vs transfer) speed up the reducers copy greatly especially when working with large data sets.

Here are some more configuration for real world cluster setup,  the docs are good too http://hadoop.apache.org/common/docs/current/cluster_setup.html

  • Some non-default configuration values used to run sort900, that is 9TB of data sorted on a cluster with 900 nodes:
    Configuration File Parameter Value Notes
    conf/hdfs-site.xml dfs.block.size 134217728 HDFS blocksize of 128MB for large file-systems.
    conf/hdfs-site.xml dfs.namenode.handler.count 40 More NameNode server threads to handle RPCs from large number of DataNodes.
    conf/mapred-site.xml mapred.reduce.parallel.copies 20 Higher number of parallel copies run by reduces to fetch outputs from very large number of maps.
    conf/mapred-site.xml mapred.child.java.opts -Xmx512M Larger heap-size for child jvms of maps/reduces.
    conf/core-site.xml fs.inmemory.size.mb 200 Larger amount of memory allocated for the in-memory file-system used to merge map-outputs at the reduces.
    conf/core-site.xml io.sort.factor 100 More streams merged at once while sorting files.
    conf/core-site.xml io.sort.mb 200 Higher memory-limit while sorting data.
    conf/core-site.xml io.file.buffer.size 131072 Size of read/write buffer used in SequenceFiles
  • Updates to some configuration values to run sort1400 and sort2000, that is 14TB of data sorted on 1400 nodes and 20TB of data sorted on 2000 nodes:
    Configuration File Parameter Value Notes
    conf/mapred-site.xml mapred.job.tracker.handler.count 60 More JobTracker server threads to handle RPCs from large number of TaskTrackers.
    conf/mapred-site.xml mapred.reduce.parallel.copies 50
    conf/mapred-site.xml tasktracker.http.threads 50 More worker threads for the TaskTracker’s http server. The http server is used by reduces to fetch intermediate map-outputs.
    conf/mapred-site.xml mapred.child.java.opts -Xmx1024M Larger heap-size for child jvms of maps/reduces.

Secondary Namenode

You may have read or heard already a dozen times that the Secondary Namenode is NOT a backup of the Namenode.  Well, let me say it again for you.  The Secondary Namenode is not the backup of the Namenode.  The Secondary Namenode should also not run on the same machine as the Namenode.  The Secondary Namenode helps to keep the edit log of the of the Namenode updated in the image file.  All of the changes that happen in HDFS are managed in memory on the Namenode.  It writes the changes to a log file that the secondary name node will compact into a new image file.  This is helpful because on restart the Namenode would have to-do this and also helpful because the logs would just grow and could be normalized.

Now, the way that the Secondary Namenode does this is with GET & POST (yes over http) to the Namenode.  It will GET the logs and POST the image.  In order for the Secondary Namenode to-do this you have to configure dfs.http.address in the hdfs-site.xml on the server that you plan to run the Secondary Namenode on.

<property>
<name>dfs.http.address</name>
<value>namenode:50070</value>
<description>
The address and the base port where the dfs namenode web ui will listen on.
If the port is 0 then the server will start on a free port.
</description>
</property>

Also, you need to put the Secondary Namenode into the masters file (which is defaulted to localhost).

Datanode Configuration Changes & Hardware

So here I only have to say that you should use JBOD (Just a Bunch Of Disks).  There have been tests showing 10%-30% degradation when HDFS blocks run on RAID.  So, when setting up your data nodes if you have a bunch of disks you can configure each one for that node to write to dfs.data.dir.

If you want you can also set a value that will always leave some space on the disk dfs.datanode.du.reserved.

Map Reduce Code Changes

In your Java code there is a little trick to help the job be “aware” within the cluster of tasks that are not dead but just working hard.  During execution of a task there is no built in reporting that the job is running as expected if it is not writing out.  So this means that if your tasks are taking up a lot of time doing work it is possible the cluster will see that task as failed (based on the mapred.task.tracker.expiry.interval setting).

Have no fear there is a way to tell cluster that your task is doing just fine.  You have 2 ways todo this you can either report the status or increment a counter.  Both of these will cause the task tracker to properly know the task is ok and this will get seen by the jobtracker in turn.  Both of these options are explained in the JavaDoc http://hadoop.apache.org/common/docs/current/api/org/apache/hadoop/mapred/Reporter.html

/*
Joe Stein
http://www.linkedin.com/in/charmalloc
*/

Categories: Cluster, Hadoop Tags:
Follow

Get every new post delivered to your Inbox.

Join 44 other followers