6. MapReduce v1 Limitations
Scalability
Maximum cluster size is 4,000 nodes and maximum concurrent tasks is 40,000
Availability
JobTracker failure kills all queued and running jobs
Resources Partitioned into Map and Reduce
Hard partitioning of Map and Reduce slots led to low resource utilization
No Support for Alternate Paradigms / Services
Only MapReduce batch jobs, nothing else
7. HADOOP 1.0
Single Use System
Batch Apps
Apache Hadoop 1.0: Single Use System
HDFS
(redundant, reliable storage)
MapReduce
(cluster resource management and data
processing)
Pig Hive
10. Store DATA in one place
YARN: Taking Hadoop Beyond Batch
Interact with that data in MULTIPLE WAYS
with Predictable Performance and Quality of Service
Applications Run Natively IN Hadoop
HDFS2
(redundant, reliable storage)
YARN
(cluster resource management)
BATCH
(MapReduce)
INTERACTIVE
(Tez)
ONLINE
(HBase)
STREAMING
(DataTorrent)
GRAPH
(Giraph)
11. Running all on the same Hadoop cluster to give
applications access to all the same source data!
YARN: Applications
MapReduce v2
Stream Processing
Master-WorkerOnline
In-Memory
Apache Storm
15. Scale
New programming models
and services
Improved cluster utilization
Agility
Backwards compatible with
MapReduce v1
Mixed workloads on the
same source of data
6 Benefits of YARN
17. Speed
Deliver interactive query through 100x
performance increases as compared to Hive
10.
Stinger: Interactive Query for Hive
SQL
Support the broadest array of SQL semantics
for analytic applications running against
Hadoop.
Scale
The only SQL interface to Hadoop designed for
queries that scale from Terabytes to Petabytes.
18. Dynamic Scaling
On-demand cluster size. Increase and decrease
the size with load.
HOYA: HBase (NoSQL) on YARN
Easier Deployment
APIs to create, start, stop and delete HBase
clusters.
Availability
Recover from Region Server loss with a new
container.
19. Machine Learning
Framework well suited for building machine
learning jobs.
Microsoft REEF
Scalable / Fault Tolerant
Makes it easy to implement scalable, fault-
tolerant runtime environments for a range of
computational models.
Maintain State
Users can build jobs that utilize data from
where it’s needed and also maintain state after
jobs are done.
Retainable
Evaluator
Execution
Framework
26. YARN: How It Works
ResourceManager
NodeManager
ApplicationMaster
NodeManager
NodeManager NodeManager
Scheduler
Container
Container Container
Client
27. YARN: Example App Deployment
ResourceManager
NodeManager
HOYA / HBase Master
NodeManager
NodeManager NodeManager
Scheduler
Region Server
Region Server Region Server
HOYA Client
28. Storm Vs. DataTorrent
Solution Matrix DataTorrent Apache Storm
Atomic Micro-batch 1 3
Events per Second Billions Thousands
Automated Parallelism 3
Dynamic Runtime Changes 3
Linear Scalability 3
State Checkpointing 3
31. Backwards Compatible
YARN is Backwards Compatible for your
existing MapReduce applications. You
can get value from it right away.
YARN: Key Take-Aways
Resource Management
YARN enables Fine Grained Resource
Management for better cluster
utilization.
One Source of Data
YARN allows you to interact with One
Source of Data in multiple ways while
maintaining Predictable Performance
and Quality of Service.
Enabling Smart People
YARN is a flexible framework that is
giving smart people and companies to
do amazing things with data.
YARN will be the de-facto distributed operating
system for Big Data
32. Storm Vs. DataTorrent - Detailed
Solution Matrix DataTorrent Apache Storm
Proprietary / Open Source O O
Support for Hadoop 1.x 1 1
Support for Hadoop 2.x 1 1
Native YARN 1 3
Dashboard 1 3
Extensible via Modules 1 1
Technical Support 1 1
Atomic Micro-batch 1 3
Events per Second Billions Thousands
Automated Parallelism 1 3
Dynamic Runtime Changes 1 3
High Availability 1 2
Prog. Languages Supported Java, Python, etc. Java, Python, etc.
Log Analysis 1 3
Site Operations 1 3
MapReduce Diagnostics 1 3
Open Source Operators Library 1 2
Open Source Application Templates 1 3
Complex Computations (DAG) 1 3
Linear Scalability 1 3
Security 1 3
CLI and Macros 1 3
Configuration Based Specification 1 3
State Checkpointing 1 3
33. Users forced to
create data system
silos for managing
mixed workloads
Developers forced
to abuse very
specific
MapReduce to fit
their use cases
The 1st Generation Of Hadoop
Hadoop
HBase
37. YARN: Why The De-Facto Distributed OS
Technology Adoption
100,000 nodes+ - 400,000 jobs - 10m compute hours daily
Enables Innovation
Smart people and companies to do amazing things to data
Financial Backing
568m+ invested in Hadoop contributing companies, nearly 400m in the
2013 alone
39. HDFS Write Data Flow
NameNode
Client
DataNode DataNode DataNode
1
2
4 5
67
3
Block Bytes
Block Bytes Block Bytes
Block Write Complete
AckAck
Ack
A
B
C
Editor's Notes
“Hadoop” has become synonymous with “Big Data” .. While Hadoop is a big part of the Big Data movement .. Hadoop itself is just a platform and the tools
The architecture of MapReduce came with it’s limitations.
Scalability
Even as specs rise on servers to accommodate more load it still couldn’t scale passed the max concurrent tasks.
Availability
JT failure kills all queued and running jobs
After restart they have to be resubmitted and start from the beginning
Unable to start from where it left off
Can be a huge problem if you have long running batch jobs
Resource partitioning of resources
Resources were broken up into distinct map and reduce slots which aren’t “fungible”. I love that word. Basically means they weren’t interchangeable.
Map slots might be “full” while reduce slots remain empty and vice-versa.
This needed to be addressed to to ensure the entire system could be used at max capacity for high utilization.
Lacks support
You were stuck using MapReduce
In Hadoop 1.0, all methods of accessing the data within the cluster were constrained to using MapReduce, Open-source Hadoop projects like Pig, Hive are built on top of MapReduce and even though they make MapReduce more accessible, they still suffer with it’s limitations.
You have seen some distributions move outside of Hadoop ecosystem, like Cloudera’s Impala, to get around the limitations of MapReduce to improve performance. But then, unfortunately, it isn’t community supported and lags behind in features because it doesn’t have the backing of the innovative open source committers.
The crazy thing is even with these limitations, 90% of the use cases Nick spoke about yesterday are based on this.
So, what has Trace3 found out on it journey through Big Data about YARN?
Well first of all, we discovered that it’s not the type of yarn that cats play with.
YARN will be the de-facto distributed operating system for Big Data and by the end of this hour you are going to see why we believe it is and why companies like Cloudera, Hortonworks and MapR are banking on this.
YARN is taking Hadoop beyond batch
YARN has solved the limitations of MapReduce v1
YARN gives you the ability to store all your data in one place and have mixed workloads working with that data and still getting predictable performance and QoS.
YARN is moving Hadoop beyond just MapReduce and Batch into Interactive, Online, Streaming, Graph, In-Memory, etc
Here are some of the apps that are making up that compute time
Hbase will be deployed on YARN
Which we will talk about more a bit later
Master-Worker applications
MapReduce has been moved out to it’s own application framework
Real-Time Streaming Analytics
This in my opinion is the most promising of the application types. I don’t want to steal my associate Rikin’s thunder, but he will be speaking a lot more in-depth around Real-Time Streaming Analytics in a session later today.
Graph Processing
YARN has enabled the ability to use iterative applications like Apache Giraph within your cluster where previously MapReduce v1 just wasn’t a viable option.
YARN is fairly new to the scene.
But that shouldn’t deter you from being confident in it.
It was conceived and architected by Yahoo!
And has gone through a very quick maturing process due to the open source community putting it through it paces
Currently YARN is running on over 100,000 nodes
Responsible for 400,000+ jobs and 10 million+ hours of compute time daily
Yes, I said 10 ‘millllllion’
So, what’s changed with YARN for it to be able to accomplish this?
YARN splits up the two major functions of the JobTracker into the ResourceManager and ApplicationMaster
Global ResourceManager
handles all of the cluster resources
Scheduler
performs its scheduling function based on the resource requirements of the applications
Per-node slave NodeManager
Responsible for launching application containers
Monitoring their resource usage
And reporting the same to the ResourceManger
Per-application ApplicationMaster
Responsible for negotiating the appropriate resource containers from the scheduler
Tracking their status
Monitoring for progress
Per-application Container running on a NodeManager
Let’s see how these all work together
Scale
YARN is no longer limited by 40000 concurrent tasks that MapReduce v1 had
Today YARN is already handling over 10 million hours of compute time on a daily basis
New Programming Models and Services
You aren’t limited to just MapReduce
If your app can benefit from a distributed operating system then you can utilize
Improved Cluster Utilization
YARN no longer has a hard partition of resources into map and reduce slots, it utilizes the resource leases aka “Containers” that aren’t limited to in functionality.
Agility
By moving MapReduce out and on top of YARN it gives customers more agility to make changes, upgrade and have different versions of their framework running so they don’t have to affect the entire cluster.
Backwards Compatible
What you are currently doing with Hadoop 1.x and MapReduce v1 will work with YARN.
Mixed workloads on the same data source
You can utilize the “data lake” architecture and run all your apps while still having perdictable performance and quality of service.
One of the projects that I’m keeping a close on is the Stinger project
Speed
100x speed increase from Hive10
SQL
Improve HiveSQL to make it more ANSI SQL-like
Scale
Ability to run queries on Terabytes to Petabytes of information
Another project to watch closely is HBase on YARN
Dynamic Scaling
Scales with usage .
As load increases,
Easier Deployment
HBase cluster deployment can be somewhat complicated, they are looking to correct that be allowing you to do it utilizing builtin API’s
Availability
When a “RegionServer” is lost, to recover it, it’s just deploying another container within the cluster.
This is a project that lays outside of my wheel-house but from what I’ve learned about it, it’s going to do amazing thing for Machine Learning.
I’ve also highlighted this project to add even more credibility to YARN by showing you a company like Microsoft is dedicating internal time and resources to build applications to run on YARN.
Previously a NameNode had one classification of storage media available to it.
Now NameNodes as of 2.3 have the ability to split up storage media available to it.
Adding awareness of storage media can allow HDFS to make better decisions about the placement of block data with input from applications.
An application can choose the distribution of replicas based on its performance and durability requirements.
NodeManger Restart allows for a restart of the NM without losing jobs .. They will continue where they left off after restart
Dynamic Resource Config - Currently containers are static .. They allocate a certain amount of proc / memory to each process. Now processes will have the ability to scale up within a container if resources are available within that NodeManager.
Jobs are submitted to the ResouceManager via a public submission protocol and go through an admission control phase during which security credentials are validated and various checks are performed.
The RM runs as a daemon on a dedicated machine, and acts as the central authority arbitrating resources for various competing applications in the cluster. Because it has a central and global view of the cluster resources, it can enforce properties such as fairness, capacity, and locality across nodes.
Accepted jobs are passed to the scheduler to be run. Once the scheduler has enough resources, the application is moved from accepted to running state. This involves allocating a resource lease
Aka as a container (bound JVM) - for the AM and spawning it on a node in the cluster. A record of accepted applications is written to persistent storage and recovered in case of RM failure.
The ApplicationMaster is the “head” of a job, managing all lifecycle aspects including dynamically increasing and decreasing resources consumption, managing the flow of execution and handling faults.
By delegating all these functions to AMs, YARN’s architecture gains a great deal of scalability, programming model flexibility, and improved upgrading/testing since multiple versions of the same framework can coexist.
The RM interacts with a special system daemon running on each node called the NodeManager (NM). Communications between RM and NMs are heartbeat based for scalability. NMs are responsible for monitorng resource availability, reporting faults, and container lifecycle management (e.g., starting, killing). The RM assembles its global view from these snapshots of NM state.
Jobs are submitted to the ResouceManager via a public submission protocol and go through an admission control phase during which security credentials are validated and various checks are performed.
The RM runs as a daemon on a dedicated machine, and acts as the central authority arbitrating resources for various competing applications in the cluster. Because it has a central and global view of the cluster resources, it can enforce properties such as fairness, capacity, and locality across nodes.
Accepted jobs are passed to the scheduler to be run. Once the scheduler has enough resources, the application is moved from accepted to running state. This involves allocating a resource lease
Aka as a container (bound JVM) - for the AM and spawning it on a node in the cluster. A record of accepted applications is written to persistent storage and recovered in case of RM failure.
The ApplicationMaster is the “head” of a job, managing all lifecycle aspects including dynamically increasing and decreasing resources consumption, managing the flow of execution and handling faults.
By delegating all these functions to AMs, YARN’s architecture gains a great deal of scalability, programming model flexibility, and improved upgrading/testing since multiple versions of the same framework can coexist.
The RM interacts with a special system daemon running on each node called the NodeManager (NM). Communications between RM and NMs are heartbeat based for scalability. NMs are responsible for monitorng resource availability, reporting faults, and container lifecycle management (e.g., starting, killing). The RM assembles its global view from these snapshots of NM state.
The Stinger project is tackling the speed portion by utilizing Apache Tez
Tez sits at the layer between MapReduce, Pig and Hive to optimize the execution of the these applications.
MapReduce Version consisted of 2 daemons / processes.
The JobTracker is a master node responsible for managing the cluster resources (map and reduce slots) and job scheduling.
The TaskTracker is a per-node agent and manages the map and reduce tasks.
Backwards Compatible
Whatever you are doing with Hadoop 1.0 and MapReduce today, will work with YARN.
Even though you don’t need all the capabilities of YARN right now, don’t hesitate to move to it and as new tools and applications become available on YARN your company will be able to utilize them.
One Source Of Data
YARN allows you to have that data lake with all of your data applications running against it.
While still maintaining predictable performance and quality of service
Resource Management
YARN accomplishes this by how it manage resources for better cluster utilization which translates to “more bang for your buck”.
Enabling Smart People
YARN is an extremely flexible framework that is giving smart people and companies the ability to do amazing things with data.
All these benefits add up to “YARN will be the de-facto distributed operating system for Big Data”
We see the innovation in Big Data happening on YARN and
We want to help you make the right choice now to avoid the headaches and costs that come along with making the wrong choice.
Before we can understand fully what YARN is solving we need to review what it’s replacing.
Hadoop 1.0
The initial design of Hadoop was focused on running massive MapReduce jobs to process web crawl.
Although It did end up evolving outside of initial use case and helped solve the “data silo problem”, it ended up creating a different issue, something called the “data system silo problem”.
Users were forced into creating data system silos due to mixed workloads.
HBase Example
Developers were forced to abuse the very specific MapReduce programming model to try to accommodate their user cases.
One of the biggest cost to a Hadoop cluster is copying data between the clusters to try to accommodate mixed workloads
PMCs are the people that give oversight for the project roadmap and provide guidance to the committers.
One thing to highlight may be that Hortonworks is a spin-off of Yahoo!
The committers are the ones who actually submit code to the project.
One thing to highlight may be that Hortonworks is a spin-off of Yahoo!
Question may arise how I can state that YARN will be the de-facto distributed operating system of Big Data. Here are the arguments for my conclusion / prediction.
HDFS Write Data Flow
1 – 7
Connect to the NN to establish block placement
Writing to the DNs
Once 1 copy of the data is placed, the client gets an acknowledgement
The first DN copies the file to the second DN
The second DN copies to the third DN
The DNs acknowledge to the other DNs the the copy has been completed
The DNs acknowledge to the other DNs the the copy has been completed
A, B and C
Once the files are written to the DNs the information about the new block is sent in the block report / heartbeat.