Apache Hadoop and Hive Dhruba Borthakur Apache Hadoop Developer

Apache Hadoop and Hive Dhruba Borthakur Apache Hadoop Developer Facebook Data Infrastructure [email protected], [email protected] Condor Week, April 22, 2009

Outline Architecture of Hadoop Distributed File System Hadoop usage at Facebook Ideas for Hadoop related research

Who Am I? Hadoop Developer – Core contributor since Hadoop’s infancy – Project Lead for Hadoop Distributed File System Facebook (Hadoop, Hive, Scribe) Yahoo! (Hadoop in Yahoo Search) Veritas (San Point Direct, Veritas File System) IBM Transarc (Andrew File System) UW Computer Science Alumni (Condor Project)

Hadoop, Why? Need to process Multi Petabyte Datasets Expensive to build reliability in each application. Nodes fail every day – Failure is expected, rather than exceptional. – The number of nodes in a cluster is not constant. Need common infrastructure – Efficient, reliable, Open Source Apache License The above goals are same as Condor, but – Workloads are IO bound and not CPU bound

Hive, Why? Need a Multi Petabyte Warehouse Files are insufficient data abstractions – Need tables, schemas, partitions, indices SQL is highly popular Need for an open data format – RDBMS have a closed data format – flexible schema Hive is a Hadoop subproject!

Hadoop & Hive History – Google GFS paper published July 2005 – Nutch uses MapReduce Feb 2006 – Becomes Lucene subproject Apr 2007 – Yahoo! on 1000-node cluster Jan 2008 – An Apache Top Level Project Jul 2008 – A 4000 node test cluster Dec 2004 Sept 2008 – Hive becomes a Hadoop subproject

Who uses Hadoop? Amazon/A9 Facebook Google IBM Joost Last.fm New York Times PowerSet Veoh Yahoo!

Commodity Hardware Typically in 2 level architecture – Nodes are commodity PCs – 30-40 nodes/rack – Uplink from rack is 3-4 gigabit – Rack-internal is 1 gigabit

Goals of HDFS Very Large Distributed File System – 10K nodes, 100 million files, 10 PB Assumes Commodity Hardware – Files are replicated to handle hardware failure – Detect failures and recovers from them Optimized for Batch Processing – Data locations exposed so that computations can move to where data resides – Provides very high aggregate bandwidth User Space, runs on heterogeneous OS

HDFS Architecture Cluster Membership NameNode e am ilen 1. f ckId, 2. Bl o es Nod Data Secondary NameNode Client 3.Read d ata Cluster Membership NameNode : Maps a file to a file-id and list of MapNodes DataNode : Maps a block-id to a physical location on disk SecondaryNameNode: Periodic merge of Transaction log DataNodes

Distributed File System Single Namespace for entire cluster Data Coherency – Write-once-read-many access model – Client can only append to existing files Files are broken up into blocks – Typically 128 MB block size – Each block replicated on multiple DataNodes Intelligent Client – Client can find location of blocks – Client accesses data directly from DataNode

NameNode Metadata Meta-data in Memory – The entire metadata is in main memory – No demand paging of meta-data Types of Metadata – List of files – List of Blocks for each file – List of DataNodes for each block – File attributes, e.g creation time, replication factor A Transaction Log – Records file creations, file deletions. etc

DataNode A Block Server – Stores data in the local file system (e.g. ext3) – Stores meta-data of a block (e.g. CRC) – Serves data and meta-data to Clients Block Report – Periodically sends a report of all existing blocks to the NameNode Facilitates Pipelining of Data – Forwards data to other specified DataNodes

Block Placement Current 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 Would like to make this policy pluggable

Data Correctness Use Checksums to validate data – Use 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

NameNode Failure A single point of failure Transaction Log stored in multiple directories – A directory on the local file system – A directory on a remote file system (NFS/CIFS) Need to develop a real HA solution

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

Rebalancer Goal: % disk full on DataNodes should be similar – – – – Usually run when new DataNodes are added Cluster is online when Rebalancer is active Rebalancer is throttled to avoid network congestion Command line tool

Hadoop Map/Reduce The Map-Reduce 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 unique -c cat file input map shuffle reduce output Natural for: – Log processing – Web search indexing – Ad-hoc queries

Hadoop at Facebook Production cluster – – – – – 4800 cores, 600 machines, 16GB per machine – April 2009 8000 cores, 1000 machines, 32 GB per machine – July 2009 4 SATA disks of 1 TB each per machine 2 level network hierarchy, 40 machines per rack Total cluster size is 2 PB, projected to be 12 PB in Q3 2009 Test cluster 800 cores, 16GB each

Data Flow Web Servers Scribe Servers Network Storage Oracle RAC Hadoop Cluster MySQL

Hadoop and Hive Usage Statistics : – – – – 15 TB uncompressed data ingested per day 55TB of compressed data scanned per day 3200 jobs on production cluster per day 80M compute minutes per day Barrier to entry is reduced: – 80 engineers have run jobs on Hadoop platform – Analysts (non-engineers) starting to use Hadoop through Hive

Ideas for Collaboration

Condor and HDFS Run Condor jobs on Hadoop File System – Create HDFS using local disk on condor nodes – Use HDFS API to find data location – Place computation close to data location Support map-reduce data abstraction model

Power Management Power Management – Major operating expense – Power down CPU’s when idle – Block placement based on access pattern Move cold data to disks that need less power Condor Green

Benchmarks Design Quantitative Benchmarks – Measure Hadoop’s fault tolerance – Measure Hive’s schema flexibility Compare above benchmark results – with RDBMS – with other grid computing engines

Job Sheduling Current state of affairs – FIFO and Fair Share scheduler – Checkpointing and parallelism tied together Topics for Research – Cycle scavenging scheduler – Separate checkpointing and parallelism – Use resource matchmaking to support heterogeneous Hadoop compute clusters – Scheduler and API for MPI workload

Commodity Networks Machines and software are commodity Networking components are not – High-end costly switches needed – Hadoop assumes hierarchical topology Design new topology based on commodity hardware

More Ideas for Research Hadoop Log Analysis – Failure prediction and root cause analysis Hadoop Data Rebalancing – Based on access patterns and load Best use of flash memory?

Summary Lots of synergy between Hadoop and Condor Let’s get the best of both worlds

Useful Links HDFS Design: – http://hadoop.apache.org/core/docs/current/hdfs design.html Hadoop API: – http://hadoop.apache.org/core/docs/current/api/ Hive: – http://hadoop.apache.org/hive/