Hortonworks HDP1, Apache Hadoop 2.0, NextGen MapReduce (YARN), HDFS Federation and the future of Hadoop with Arun C. Murthy
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:
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.
- 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:
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.
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.
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.
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 has two main layers:
- 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.
- Block Management (which is done in Namenode)
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.
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.
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.
- 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.
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
. 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.