Running Non-MapReduce Big Data Applications on Apache Hadoop

Editor's Notes

• #5: Jobtracker and Tasktrackers with 4 map slots and 2 reduce slots each As a job is submitted, the jobtracker runs work in assigned slots.As more jobs run, more and more slots in the cluster get used up.
• #6: Talk about a MR job structure – input, map, sort, disk, shuffle, merge, reduce, output. Inspite of all these restrictions, the hadoop ecosystem has grown coming up with innovative ways to do things.
• #9: YARN takes what was a single-use data platform for batch processing and evolving it into a multi-use platform that enables batch, interactive, real-time and stream processing.
• #11: YARN acts as the primary resource management layer and mediator of access to the cluster, giving anyonethe capability to store data in a single place and then interact with it in multiple ways, simultaneously, with consistent levels of service.
• #13: Now, compared to v1, a YARN base cluster has the RM as the central controlling entity and NodeManagers on all the worker nodes. There is no notion of slots ( maps/reduces ). There is only controlled resources. Memory/CPU, disks/network IO in the future. Explain application flow – client submits application, RM launches Application Master. App Master requests containers from the RM and then launches containers. Explain how subsequent applications work – containers launched as long as they fit within the resource pool.
• #15: YarnClientApplication Lifecycle Control – submit, query state, kill - .e.g Submit, followed by poll state, and optionally connect to AM for further stateCluster Information – Query node status, Scheduler queues etcAMRMCLientRequest Resources from the RM – specifies a priority (within app only), location information, and whether strict locality is required, also number of such containers.Hides the details of the protocol – which is absolute (i.e. each call to the RM must provide complete information about a priority level)Once a container is allocated, provides a utility method to link it back to the entity which generated the request (entity optionally passed in while asking for containers)Allocate call doubles as a heartbeat – must be sent out ever so often.Also provides some node level information, available resources, etcNMClientOnce resources are obtained from the RM, an App uses this to communicate with the NM to start containers. Setup the ContainerLuanchContext – cmdLine, environment, resources.Authorization information is part of the contianers allocated by the RM – to prevent unauthorized usageOnly a single process can be launched on an allocated container, and one must be within a certain period otherwise the allocation times outContainer must launch a process with a certain time after being allocatedOnly a single process can be launched on an allocated container
• #16: Covered YARN Architecture, What a YARN application looks like, and Haddop 1vsHadoop 2
• #17: Application functionality – stream processing, batch, interactive, etcRest – Application’s interaction with YARNShort running – obtain, process, releaseLong Running – obtain, keep, scale up and down if required.How would we categorize applications? YARN is still nascent. We will see more and more applications being built on it. Today, based on the current ecosystem, there are various ways of categorizing applications. Data-processing execution applications such as MapReduce and TezStreaming applications like Samza, StormServices like Hbase on YARNOn demand web server scalingFrom a framework perspective or from the perspective of building applications, they can be categorized according to How long the application runs, How the application accepts work, Language it’s written in.Application Lifetime‘Short’ Running Jobs – expected to finish their processing and complete. Typically one YARN application per job. - Ask for adequate resources, process and then release the resources once the job completes.Long Running Services – typically deployed via YARN to share cluster resources. Will generally reserve certain resources on start-up. During the lifetime of the job, new resources can be acquired – and it’s up to the application on whether it wants to hang on to these or return them to the cluster. May define their own job/work submission construct – e.g. Storm starts up the Storm master which then accepts work – each of these is not a separate YARN application
• #18: Hadoop associated with JavaSimpler to write non Java apps – but not trivialImplement RPC mechanics in language of choice, and then the HDFS/YARN protocolsOR … listed optionsEveryone associates Hadoop with Java. Hadoop RPC  while that’s primarily the case, writing non Java applications should be easier in Hadoop 2Either build an RPC layer in the language being used – which can then be used to access the HDFS and YARN protocols (disclaimer: may not be a trivial exercise)HDFS –WebHDFS which is a REST based API, libHdfs which is JNI based C API for accessing HDFS. Folks at Spotify labs have built a Phthon library using Hadoop RPC mechanics – also have detailed out the RPC interactionsYARN – Experimental early stage Go RPC by Arun Murthy – note: this is experimental and very early stage. - Alternately, write the Applcation Logic in Java – but allow the rest of the application to be in any languageIf a language binding exists, barring some caveats, you should be able to write yarn applications. Recently, Arun Murthy came up with a hello world application built in GO. Please do take this with a bucket of salt as there are still some things such tokens and security which need to flush out properly but effectively, soon, we should have applications coming up that in written in other languages apart from Java.
• #19: Computation expressed as a DAGMR limited to two stages Tez allows multiple stages with different types of dataflow between stagesLaunch AM, StartContaienr – specify command line, environment, resources etc.
• #20: Deploy Hbase on demand - Old: Separate node or carve out part of each node.e.g. If Hbase not always required – hourly / daily jobs – spin up on demand, or running custom co-processors, test clustersStoresHbase cluster stateCLI tool for all functionality.Locality, dynamic resizing, failure handlingCLI Tool + configuration to deployHBase clusters on YARNRather than carving out separate nodes, or resources on each node up-front – allows the cluster resources to be centrally managed by YARN.Hbase Master starts up directly within the HOYA ApplicationMaster.Region Servers are run within containers allocated by YARN – based on previously stored state or Locality.HOYA CLI tool used to manage the cluster – restart, stop, resize etc.Example usage: Bring up the cluster only when it’s required, and then release resources back once it’s complete – a daily analysis job, isolation between users ?Resources for the cluster are obtained from YARN – allowing it to share resources with the rest of the cluster. e.g. Daily analysis job, bring up a previously persisted HBase cluster, execute and release resources.Provides APIs to grow / shrink a clusterHOYA AM notified by YARN on container or node failure – at which point it can bring up a replacement RegionServer on another node
• #21: Stream processing –YARN resource scheduling, rather than starting daemons – which also know how to handle resources.CGroups, Samza – A stream processing system built to make use of YARN as the execution framework - no separate daemon for scheduling resources.Samza AM spun up per job / dataflow.Takes care of making node local requests for Samza ContainersRe-scheduling failed tasks – new container from YARN
• #22: Graph processing – Giraph, HamaStream proessing – Smaza, Storm, Spark, DataTorrentMapReduceTez – fast query executionWeave/REEF – frameworks to help with writing applicationsList of some of the applications which already support YARN, in some form.Smaza, Storm, S4 and DataTorrent are streaming frameworksVarious types of graph processing frameworks – Giraph and Hama are graph processing systemsThere’s some github projects – caching systems, on-demand web-server spin up Wave and REEF are frameworks on top of YARN to make writing applications easier
• #23: Recommendations / Points to keep in mind
• #24: Client libs – mostly thin, AMRM hides detailsResource Negotiation – contending withh other appsFailure handling – large cluster, failed node, disk, network etcCheckpointing – depends on apps – long running should maintain AM state, maybe stask stateUse Provided Client libraries In most part, these are thin wrappers over the YARN protocols – but in some cases do have some additional functionalityResource Negotiation Shared cluster, app may be contending with others. Write your application to work with incremental resource allocations; Be aware of this if the application MUST have certain resources Locality requests are best effort, unless strict locality is enabled – delays before rack or non, local but no guaranteesFailure Handling Design applications to manage node failure, container failure etc. YARN will restart AMs, and will notify the AM about Container failure – it’s up to the application to deal with this, which brings us to CheckpointingCheckpointing In case of failure or restart, the AM will be restarted by YARN – recovered AMs can be written to restart from scratch, or can store state to recover from where they left off. Similarly – the AM may need to store task state. As an example – MapReduce stores task state in HDFS, on AM restart – it’s able to recover successful work from the previous run and continue the job from there.
• #25: Trying to avoid complicated cluster deploysDeploy everything with the AppClient libraries – debuggingSecurity – Secure cluster – YARN interaction uses kerberos / tokens. Client to AM / Container to AM up to the AppCluster Dependencies Application can run on any node in the cluster. Be aware of what’s available on the cluster, or ship all resources and binaries required by an application along with it. Typically, a shell, Java can be assumed.Client Only Install Recommended to deploy applications via a client package – YARN will take care of making this available on the nodes where the application runs. Allows for different versions of the same application to co-exist on the cluster.Also simplifies the initial cluster deployment. MapReduce is currently deployed on all nodes – but is moving to a model where it’ll only be deployed on the client. Tez, HOYA etc are already client side only applications.Quick note on Security YARN, in secure mode, use kerberos and tokens to secure RM and NM communication. Communication with the AM and between containers needs to be secure by the app-writer. Hadoop RPC security can ofcourse be used.
• #26: MiniYARN – cluster within process in threads (HDFS, MR, Hbase) – avoid for unit tests – SLOWUnmanagedAM – debugger, heap dumps, etcLogs – enable by default – All application logsMiniYarnCluster Those familiar with MapReduce, HDFS, HBase – this is a YARN cluster which spins up within the same JVM. Can configure number of NodeManagers etc. Primarily meant for regression testing – should be avoided for unit tests considering the cost.Unmanaged AM Allows the AM to be run outside the cluster. Typically, when running on a cluster – the AM could run on any node which can make it tough to debug. Helps with attaching debuggers, accessing AM logs etc.Log Aggregation – aggregates logs onto HDFS Recommend running with log aggregation enabled – enormously useful for debugging. Can fetch ALL logs for an application after it completes, or logs for specific containers with a simple CLI command.
• #27: HA and work preserving – being actively worked upoin by the communitiy.Scheduler – Additional resources – specifically disk / network. Gang schedulingRolling upgrades – upgrading a cluster typically involves downtime. NM forgets containers across restartsLong Running – Enhandcement to log handling, security, multiple tasks per container, container resizingHA and work preserving restart are still being worked on in the community – YARN-128 and YARN-149.On scheduling – there’ve been requests for gang scheduling, meeting SLAs. Also TBD is support for scheduling additional resource types – disk/ network.Rolling Upgrades – some work pending. Big piece here, which ties in with work preserving restart – restarting a NodeManager should not cause processes started by the previous NM to be killedLong Running Services support – handling logs, security – specifically token expiryAdditional utility libraries to help AppWriters – primarily geared towards checkpointing in the AM, app history handling