Hadoop, a distributed framework for Big Data Class: CS 237

Hadoop, a distributed framework for Big Data Class: CS 237 Distributed Systems Middleware Instructor: Nalini Venkatasubramanian

Introduction 1. Introduction: Hadoop’s history and advantages 2. Architecture in detail 3. Hadoop in industry

What is Hadoop? Apache top level project, open-source implementation of frameworks for reliable, scalable, distributed computing and data storage. It is a flexible and highly-available architecture for large scale computation and data processing on a network of commodity hardware.

Brief History of Hadoop Designed to answer the question: “How to process big data with reasonable cost and time?”

Search engines in 1990s 1996 1996 1996 1997

Google search engines 1998 2013

Hadoop’s Developers Doug Cutting 2005: Doug Cutting and Michael J. Cafarella developed Hadoop to support distribution for the Nutch search engine project. The project was funded by Yahoo. 2006: Yahoo gave the project to Apache Software Foundation.

Google Origins 2003 2004 2006

Some Hadoop Milestones 2008 - Hadoop Wins Terabyte Sort Benchmark (sorted 1 terabyte of data in 209 seconds, compared to previous record of 297 seconds) 2009 - Avro and Chukwa became new members of Hadoop Framework family 2010 - Hadoop's Hbase, Hive and Pig subprojects completed, adding more computational power to Hadoop framework 2011 - ZooKeeper Completed 2013 - Hadoop 1.1.2 and Hadoop 2.0.3 alpha. - Ambari, Cassandra, Mahout have been added

What is Hadoop? Hadoop: an open-source software framework that supports dataintensive distributed applications, licensed under the Apache v2 license. Goals / Requirements: Abstract and facilitate the storage and processing of large and/or rapidly growing data sets Structured and non-structured data Simple programming models High scalability and availability Use commodity (cheap!) hardware with little redundancy Fault-tolerance Move computation rather than data

Hadoop Framework Tools

Hadoop’s Architecture Distributed, with some centralization Main nodes of cluster are where most of the computational power and storage of the system lies Main nodes run TaskTracker to accept and reply to MapReduce tasks, and also DataNode to store needed blocks closely as possible Central control node runs NameNode to keep track of HDFS directories & files, and JobTracker to dispatch compute tasks to TaskTracker Written in Java, also supports Python and Ruby

Hadoop’s Architecture

Hadoop’s Architecture Hadoop Distributed Filesystem Tailored to needs of MapReduce Targeted towards many reads of filestreams Writes are more costly High degree of data replication (3x by default) No need for RAID on normal nodes Large blocksize (64MB) Location awareness of DataNodes in network

Hadoop’s Architecture NameNode: Stores metadata for the files, like the directory structure of a typical FS. The server holding the NameNode instance is quite crucial, as there is only one. Transaction log for file deletes/adds, etc. Does not use transactions for whole blocks or file-streams, only metadata. Handles creation of more replica blocks when necessary after a DataNode failure

Hadoop’s Architecture DataNode: Stores the actual data in HDFS Can run on any underlying filesystem (ext3/4, NTFS, etc) Notifies NameNode of what blocks it has NameNode replicates blocks 2x in local rack, 1x elsewhere

Hadoop’s Architecture: MapReduce Engine

Hadoop’s Architecture MapReduce Engine: JobTracker & TaskTracker JobTracker splits up data into smaller tasks(“Map”) and sends it to the TaskTracker process in each node TaskTracker reports back to the JobTracker node and reports on job progress, sends data (“Reduce”) or requests new jobs

Hadoop’s Architecture None of these components are necessarily limited to using HDFS Many other distributed file-systems with quite different architectures work Many other software packages besides Hadoop's MapReduce platform make use of HDFS

Hadoop in the Wild Hadoop is in use at most organizations that handle big data: o Yahoo! o Facebook o Amazon o Netflix o Etc Some examples of scale: o Yahoo!’s Search Webmap runs on 10,000 core Linux cluster and powers Yahoo! Web search o FB’s Hadoop cluster hosts 100 PB of data (July, 2012) & growing at ½ PB/day (Nov, 2012)

Hadoop in the Wild Three main applications of Hadoop: Advertisement (Mining user behavior to generate recommendations) Searches (group related documents) Security (search for uncommon patterns)

Hadoop in the Wild Non-realtime large dataset computing: o NY Times was dynamically generating PDFs of articles from 1851-1922 o Wanted to pre-generate & statically serve articles to improve performance o Using Hadoop MapReduce running on EC2 / S3, converted 4TB of TIFFs into 11 million PDF articles in 24 hrs

Hadoop in the Wild: Facebook Messages Design requirements: o Integrate display of email, SMS and chat messages between pairs and groups of users o Strong control over who users receive messages from o Suited for production use between 500 million people immediately after launch o Stringent latency & uptime requirements

Hadoop in the Wild System requirements o High write throughput o Cheap, elastic storage o Low latency o High consistency (within a single data center good enough) o Disk-efficient sequential and random read performance

Hadoop in the Wild Classic alternatives o These requirements typically met using large MySQL cluster & caching tiers using Memcached o Content on HDFS could be loaded into MySQL or Memcached if needed by web tier Problems with previous solutions o MySQL has low random write throughput BIG problem for messaging! o Difficult to scale MySQL clusters rapidly while maintaining performance o MySQL clusters have high management overhead, require more expensive hardware

Hadoop in the Wild Facebook’s solution o Hadoop HBase as foundations o Improve & adapt HDFS and HBase to scale to FB’s workload and operational considerations Major concern was availability: NameNode is SPOF & failover times are at least 20 minutes Proprietary “AvatarNode”: eliminates SPOF, makes HDFS safe to deploy even with 24/7 uptime requirement Performance improvements for realtime workload: RPC timeout. Rather fail fast and try a different DataNode

Hadoop Highlights Distributed File System Fault Tolerance Open Data Format Flexible Schema Queryable Database

Why use Hadoop? Need to process Multi Petabyte Datasets Data may not have strict schema Expensive to build reliability in each application Nodes fails everyday Need common infrastructure Very Large Distributed File System Assumes Commodity Hardware Optimized for Batch Processing Runs on heterogeneous OS

DataNode A Block Sever – Stores data in local file system – Stores meta-data of a block - checksum – Serves data and meta-data to clients Block Report – Periodically sends a report of all existing blocks to NameNode Facilitate Pipelining of Data – Forwards data to other specified DataNodes

Block Placement Replication Strategy – One replica on local node – Second replica on a remote rack – Third replica on same remote rack – Additional replicas are randomly placed Clients read from nearest replica

Data Correctness Use Checksums to validate data – CRC32 File Creation – Client computes checksum per 512 byte – DataNode stores the checksum File Access – Client retrieves the data and checksum from DataNode – If validation fails, client tries other replicas

Data Pipelining Client retrieves a list of DataNodes on which to place replicas of a block Client writes block to the first DataNode The first DataNode forwards the data to the next DataNode in the Pipeline When all replicas are written, the client moves on to write the next block in file

Hadoop MapReduce MapReduce programming model – Framework for distributed processing of large data sets – Pluggable user code runs in generic framework Common design pattern in data processing – cat * grep sort uniq -c cat file – input map shuffle reduce output

MapReduce Usage Log processing Web search indexing Ad-hoc queries

Closer Look MapReduce Component – JobClient – JobTracker – TaskTracker – Child Job Creation/Execution Process

MapReduce Process (org.apache.hadoop.mapred) JobClient – Submit job JobTracker – Manage and schedule job, split job into tasks TaskTracker – Start and monitor the task execution Child – The process that really execute the task

Inter Process Communication IPC/RPC (org.apache.hadoop.ipc) Protocol JobSubmissionProtocol – JobClient ------------- JobTracker InterTrackerProtocol – TaskTracker ------------ JobTracker TaskUmbilicalProtocol – TaskTracker ------------- Child JobTracker impliments both protocol and works as server in both IPC TaskTracker implements the TaskUmbilicalProtocol; Child gets task information and reports task status through it.

JobClient.submitJob - 1 Check input and output, e.g. check if the output directory is already existing – job.getInputFormat().validateInput(job); – job.getOutputFormat().checkOutputSpecs(fs, job); Get InputSplits, sort, and write output to HDFS – InputSplit[] splits job.getInputFormat(). getSplits(job, job.getNumMapTasks()); – writeSplitsFile(splits, out); // out is SYSTEMDIR/ JOBID/job.split

JobClient.submitJob - 2 The jar file and configuration file will be uploaded to HDFS system directory – job.write(out); // out is SYSTEMDIR/ JOBID/job.xml JobStatus status jobSubmitClient.submitJob(jobId); – This is an RPC invocation, jobSubmitClient is a proxy created in the initialization

Job initialization on JobTracker - 1 JobTracker.submitJob(jobID) -- receive RPC invocation request JobInProgress job new JobInProgress(jobId, this, this.conf) Add the job into Job Queue – jobs.put(job.getProfile().getJobId(), job); – jobsByPriority.add(job); – jobInitQueue.add(job);

Job initialization on JobTracker - 2 Sort by priority – resortPriority(); – compare the JobPrioity first, then compare the JobSubmissionTime Wake JobInitThread – jobInitQueue.notifyall(); – job jobInitQueue.remove(0); – job.initTasks();

JobInProgress - 1 JobInProgress(String jobid, JobTracker jobtracker, JobConf default conf); JobInProgress.initTasks() – DataInputStream splitFile fs.open(new Path(conf.get(“mapred.job.split.file”))); // mapred.job.split.file -- SYSTEMDIR/ JOBID/job.split

JobInProgress - 2 splits JobClient.readSplitFile(splitFile); numMapTasks splits.length; maps[i] new TaskInProgress(jobId, jobFile, splits[i], jobtracker, conf, this, i); reduces[i] new TaskInProgress(jobId, jobFile, splits[i], jobtracker, conf, this, i); JobStatus -- JobStatus.RUNNING

JobTracker Task Scheduling - 1 Task getNewTaskForTaskTracker(String taskTracker) Compute the maximum tasks that can be running on taskTracker – int maxCurrentMap Tasks tts.getMaxMapTasks(); – int maxMapLoad Math.min(maxCurrentMapTasks, (int)Math.ceil(double) remainingMapLoad/numTaskTrackers));

JobTracker Task Scheduling - 2 int numMaps tts.countMapTasks(); // running tasks number If numMaps maxMapLoad, then more tasks can be allocated, then based on priority, pick the first job from the jobsByPriority Queue, create a task, and return to TaskTracker – Task t job.obtainNewMapTask(tts, numTaskTrackers);

Start TaskTracker - 1 initialize() – Remove original local directory – RPC initialization TaskReportServer RPC.getServer(this, bindAddress, tmpPort, max, false, this, fConf); InterTrackerProtocol jobClient (InterTrackerProtocol) RPC.waitForProxy(InterTrackerProtocol.class, InterTrackerProtocol.versionID, jobTrackAddr, this.fConf);

Start TaskTracker - 2 run(); offerService(); TaskTracker talks to JobTracker with HeartBeat message periodically – HeatbeatResponse heartbeatResponse transmitHeartBeat();

TaskTracker.localizeJob(TaskInProgress tip); Run Task on TaskTracker - 1 launchTasksForJob(tip, new JobConf(rjob.jobFile)); – tip.launchTask(); // TaskTracker.TaskInProgress – tip.localizeTask(task); // create folder, symbol link – runner task.createRunner(TaskTracker.this); – runner.start(); // start TaskRunner thread

TaskRunner.run(); Run Task on TaskTracker - 2 – Configure child process’ jvm parameters, i.e. classpath, taskid, taskReportServer’s address & port – Start Child Process runChild(wrappedCommand, workDir, taskid);

Child.main() Create RPC Proxy, and execute RPC invocation – TaskUmbilicalProtocol umbilical (TaskUmbilicalProtocol) RPC.getProxy(TaskUmbilicalProtocol.class, TaskUmbilicalProtocol.versionID, address, defaultConf); – Task task umbilical.getTask(taskid); task.run(); // mapTask / reduceTask.run

Finish Job - 1 Child – task.done(umilical); RPC call: umbilical.done(taskId, shouldBePromoted) TaskTracker – done(taskId, shouldPromote) TaskInProgress tip tasks.get(taskid); tip.reportDone(shouldPromote); – taskStatus.setRunState(TaskStatus.State.SUCCEEDED)

Finish Job - 2 JobTracker – TaskStatus report: status.getTaskReports(); – TaskInProgress tip taskidToTIPMap.get(taskId); – JobInProgress update JobStatus tip.getJob().updateTaskStatus(tip, report, myMetrics); – One task of current job is finished – completedTask(tip, taskStatus, metrics); – If (this.status.getRunState() JobStatus.RUNNING && allDone) {this.status.setRunState(JobStatus.SUCCEEDED)}

Demo Word Count – hadoop jar hadoop-0.20.2-examples.jar wordcount input dir output dir Hive – hive -f pagerank.hive