Etcetera! CMSC 491 Hadoop-Based Distributed Computing Spring 2015 Adam Shook

Transcription

1 Etcetera! CMSC 491 Hadoop-Based Distributed Computing Spring 2015 Adam Shook

2 Agenda Advanced HDFS Features Apache Kafka Apache Cassandra Redis (but more this time) Cluster Planning

3 ADVANCED HDFS FEATURES

4 Highly Available NameNode Highly Available NameNode feature eliminates SPOF Requires two NameNodes and some extra configuration Active/Passive or Active/Active Clients only contact the active NameNode DataNodes report in and heartbeat with both NameNodes Active NameNode writes metadata to a quorum of JournalNodes Standby NameNode reads the JournalNodes to stay in sync There is no CheckPointNode (SecondaryNameNode) The passive NameNode performs checkpoint operations

5 HA NameNode Failover There are two failover scenarios Graceful Performed by an administrator for maintenance Automated Active NameNode fails Failed NameNode must be fenced Eliminates the 'split brain syndrome' Two fencing methods are available sshfence Kill NameNodes daemon shell script disables access to the NameNode, shuts down the network switch port, sends power off to the failed NameNode There is no 'default' fencing method

6 Release lock ZKFC NN Active ZooKeeper NFS or QJM Shared NN State Lock Released Created Create Lock ZKFC NN Become Active Standby Active I'm the Boss Data Node Data Node Data Node

7 HDFS Federation Useful for: Isolation/multi-tenancy Horizontal scalability of HDFS namespace Performance Allows for multiple independent NameNodes using the same collection of DataNodes DataNodes store blocks from all NameNode pools

8 Federated NameNodes File-system namespace scalable beyond heap size NameNode performance no longer a bottleneck NameNode failure/degradation is isolated Only data managed by the failed NameNode is unavailable Each NameNode can be made Highly Available

9 Hadoop Security Hadoop's original design web crawler and indexing Not designed for processing of confidential data Small number of trusted users Access to cluster controlled by providing user accounts Little / no control on what a user could do once logged in HDFS permissions were added in the Hadoop 0.16 release Similar to basic UNIX file permissions HDFS permissions can be disabled via dfs.permissions Basically for protection against user-induced accidents Did not protect from attacks Authentication is accomplished on the client side Easily subverted via a simple configuration parameter

10 Kerberos Kerberos support introduced in the Hadoop release Developed at MIT / freely available Not a Hadoop-specific feature Not included in Hadoop releases Works on the basis of 'tickets' Allow communicating nodes to securely identify each other across unsecure networks Primarily a client/server model implementing mutual authentication The user and the server verify each other's identity

11 How Kerberos Works Client forwards the username to KDC A. KDC sends Client/TGS Session Key, encrypted with user's password B. KDC issues a TGT, encrypted with TGS's key C. Sends B and service ID to TGS D. Authenticator encrypted w/a E. TGS issues CTS ticket, encrypted with SS key F. TGS issues CSS, encrypted w/a G. New authenticator encrypted with F H. Timestamp found in G+1 KDC - Key Distribution Center TGS Ticket Granting Service TGT Ticket Granting Ticket CTS Client-to-Server Ticket CSS Client Server Session Key

12 Kerberos Services Authentication Server Authenticates client Gives client enough information to authenticate with Service Server Service Server Authenticates client Authenticates itself to client Provides services to client

13 Kerberos Limitations Single point of failure Must use multiple servers Implement failback authentication mechanisms Strict time requirements 'tickets' are time stamped Clocks on all host must be carefully synchronized All authentication is controlled by the KDC Compromise of this infrastructure will allow attackers to impersonate any user Each network service requiring a different host name must have its own set of Kerberos keys Complicates virtual hosting of clusters

14 APACHE KAFKA

15 Overview Kafka is a publish-subscribe messaging rethought as a distributed commit log Fast Scalable Durable Distributed

16 Kafka adoption and use cases LinkedIn: activity streams, operational metrics, data bus 400 nodes, 18k topics, 220B msg/day (peak 3.2M msg/s), May 2014 Netflix: real-time monitoring and event processing Twitter: as part of their Storm real-time data pipelines Spotify: log delivery (from 4h down to 10s), Hadoop Loggly: log collection and processing Mozilla: telemetry data Airbnb, Cisco, Gnip, InfoChimps, Ooyala, Square, Uber, 16

17 How fast is Kafka? Up to 2 million writes/sec on 3 cheap machines Using 3 producers on 3 different machines, 3x async replication Only 1 producer/machine because NIC already saturated Sustained throughput as stored data grows Slightly different test config than 2M writes/sec above. 17

18 Why is Kafka so fast? Fast writes: While Kafka persists all data to disk, essentially all writes go to the page cache of OS, i.e. RAM. Fast reads: Very efficient to transfer data from page cache to a network socket Linux: sendfile() system call Combination of the two = fast Kafka! Example (Operations): On a Kafka cluster where the consumers are mostly caught up you will see no read activity on the disks as they will be serving data entirely from cache. 18

19 A first look The who is who Producers write data to brokers. Consumers read data from brokers. All this is distributed. The data Data is stored in topics. Topics are split into partitions, which are replicated. 19

20 A first look 20

21 Topics Topic: feed name to which messages are published Example: zerg.hydra Kafka prunes head based on age or max size or key Kafka topic new Producer A1 Producer A2 Producer An Older msgs Newer msgs Producers always append to tail (think: append to a file) Broker(s) 21

22 Topics Consumer group C1 Consumer group C2 Consumers use an offset pointer to track/control their read progress (and decide the pace of consumption) Older msgs Newer msgs new Producer A1 Producer A2 Producer An Producers always append to tail (think: append to a file) Broker(s) 22

23 A topic consists of partitions. Partitions Partition: ordered + immutable sequence of messages that is continually appended to 23

24 Partitions #partitions of a topic is configurable #partitions determines max consumer (group) parallelism cf. parallelism of Storm s KafkaSpout via builder.setspout(,,n) Consumer group A, with 2 consumers, reads from a 4-partition topic Consumer group B, with 4 consumers, reads from the same topic 24

25 Partition offsets Offset: messages in the partitions are each assigned a unique (per partition) and sequential id called the offset Consumers track their pointers via (offset, partition, topic) tuples Consumer group C1 25

26 Replicas of a partition Replicas: backups of a partition They exist solely to prevent data loss. Replicas are never read from, never written to. They do NOT help to increase producer or consumer parallelism! Kafka tolerates (numreplicas - 1) dead brokers before losing data LinkedIn: numreplicas == 2 1 broker can die 26

27 APACHE CASSANDRA

28 In a couple dozen words... Apache Cassandra is an open source, distributed, decentralized, elastically scalable, highly available, fault-tolerant, tunably consistent, column-oriented database with a lot of adjectives

29 Overview Originally created by Facebook and opened sourced in 2008 Based on Google Big Table & Amazon Dynamo Massively Scalable Easy to use No relation to Hadoop Specifically, data is not stored on HDFS

30 Distributed and Decentralized Distributed Can run on multiple machines Decentralized No single point of failure No master or slave issues by using a peer-to-peer architecture (gossip protocol, specifically) Can run across geographic datacenters

31 Elastic Scalability Scales horizontally Adding nodes linearly increases performance Decreasing and increasing nodecounts happen seamlessly

32 Highly Available and Fault Tolerant Multiple networked computers in a cluster Facility for recognizing node failures Forward failing over requests to another part of the system

33 Tunable Consistency Choice between strong and eventual consistency Adjustable for reads and write operations separately Conflicts are solved during reads

34 Stored in spare multidimensional hash tables Row can have multiple columns, and not necessarily the same amount of columns for each row Each row has a unique key used for partitioning Column-Oriented

35 Query with CQL Familiar SQL-like syntax that maps to Cassandra's storage engine and simplifies data modeling CREATE TABLE songs ( id uuid PRIMARY KEY, title text, album text, artist text, data blob, tags set <text> ); INSERT INTO songs (id, title, artist, album, tags) VALUES ( 'a3e648f...', 'La Grange', 'ZZ Top', 'Tres Hombres', {'cool', 'hot'}); SELECT * FROM songs WHERE id = 'a3e648f...';

36 When should I use this? Key features to compliment a Hadoop system: Geographical distribution Large deployments of structured data

37 REDIS

38 Introduction ANSI C open-source advanced key-value store Commonly referred to as a data structure server, since keys can contain strings, hashes, lists, sets, and sorted sets Operations are atomic and there are a bunch of them All data is stored in-memory, and can be persisted using snapshots or transaction logs Trivial master-slave replication

39 Clients Redis itself is ANSI C, but the protocol is opensource and developers have created support in many languages C C# C++ Clojure Common Lisp D Dart Emacs lisp Erland Fancy GNU Prolog Go Haskell haxe Java Lua Node.js Objective-C Perl PHP Pure Data Python Ruby Rust Scala Scheme Smalltalk Tcl

40 Data Types Redis keys can be anything from a string to a byte array of a JPEG Keys have associated data types, and we should talk about them Strings Lists Hashes Sets Sorted Sets HyperLogLogs

41 Strings! The simplest type Supports a number of operations, including sets, gets, and incremental operations for values > SET mkey "my binary safe value" OK > GET mkey "my binary safe value"

42 Lists! Linked Lists, actually, i.e. O(1) for inserts into the head or tail of the list Accessing an element by index... O(N) > RPUSH messages "Hello how are you?: (integer) 1 > RPUSH messages "Fine thanks. I'm having fun with Redis" (integer) 2 > RPUSH messages "I should look into this NOSQL thing ASAP" (integer) 3 > LRANGE messages 0 2 1) "Hello how are you?" 2) "Fine thanks. I'm having fun with Redis" 3) "I should look into this NOSQL thing ASAP"

43 Hashes! Maps between string fields and string values > HMSET user:1000 username antirez password P1pp0 age 34 OK > HGETALL user:1000 1) "username" 2) "antirez" 3) "password" 4) "P1pp0" 5) "age" 6) "34" > HSET user:100 password (integer) 0 > HGETALL user:1000 1) "username" 2) "antirez" 3) "password" 4) "12345" 5) "age" 6) "34"

44 Sets! Unordered collection of strings Supports adds, gets, is-member checks, intersections, unions, sorting... > SADD myset 1 (integer) 1 > SADD myset 2 (integer) 1 > SADD myset 3 (integer) 1 > SMEMBERS myset 1) "1" 2) "2" 3) "3" > SADD myotherset 2 (integer) 1 > SINTER myset myotherset 1) "2" > SUNION myset myotherset 1) "1" 2) "2" 3) "3"

45 Sorted Sets! Similar to sorted sets, but they have an associated score and can return items in order Elements are already sorted via an O(log(n)) operation, so returning them is easy > ZADD hackers 1940 "Alan Kay" > ZRANGE hackers 0-1 (integer) 1 1) "Alan Turing" > ZADD hackers 1953 "Richard Stallman" 2) "Claude Shannon" (integer) 1 3) "Alan Kay" > ZADD hackers 1965 "Yukihiro Matsumoto" 4) "Richard Stallman" (integer) 1 5) "Yukihiro Matsumoto" > ZADD hackers 1916 "Claude Shannon" 6) "Linus Torvalds" (integer) 1 > ZADD hackers 1969 "Linus Torvalds" (integer) 1 > ZADD hackers 1912 "Alan Turing" (integer) 1

46 HyperLogLogs! Probabilistic data structure to estimate the cardinality of a set Very useful when you have a set with high cardinality Talking millions Returns 1 if the cardinality changed, 0 otherwise > PFADD hll a b c d e f g (integer) 1 > PFCOUNT hll (integer) 7 > PFADD hll a (integer) 0 > PFADD hll h (integer) 1 > PFCOUNT hll (integer) 8

47 Features Transactions Pub/Sub Lua Scripting Key Expiration Redis Clustering

48 Transactions Guarantees no client requests are served in the middle of a transaction Either all commands or none are processed, so they are atomic MULTI begins a transaction, and EXEC commits it Redis will queue commands and process them upon EXEC All commands in the queue are processed, even if one fails > MULTI OK > INCR foo QUEUED > INCR bar QUEUED > EXEC 1) (integer) 1 2) (integer) 1

49 Pub/Sub Messaging paradigm where publishers send messages to subscribers (if any) via channels Subscribers express interest in channels, and receive messages from publishers (if any) SUBSCRIBE test Clients can subscribe to channels and messages from publishers will be pushed to them by Redis PUBLISH test Hello Can do pattern-based subscriptions to channels PSUBSCRIBE news.*

50 Lua Scripting You can run Lui scripts to manipulate Redis > eval "return redis.call('set','foo','bar')" 0 OK

51 Expire Keys after time Set a timeout on a key, having Redis automatically delete it after the set time Use case: Maintain session information for a user for the last 60 seconds to recommend related products MULTI RPUSH pagewviews.user:<userid> EXPIRE pagewviews.user:<userid> 60 EXEC

52 Redis Cluster Redis Cluster is not production ready, but can be used to do partitioning of your data cross multiple Redis instances A few abstractions exist today to partition among Multiple instances, but they are not out-of-the-box with a Redis download

53 Use Cases Session Cache Ranking lists Auto Complete Twitter/Github/Pinterest/Snapchat/Craiglist/ StackOverflow/Flicker

54 CLUSTER PLANNING

55 Workload Considerations Balanced workloads Jobs are distributed across various job types CPU bound Disk I/O bound Network I/O bound Compute intensive workloads - Data Analytics CPU bound workloads require: Large numbers of CPU's Large amounts of memory to store in-process data I/O intensive workloads - Sorting I/O bound workloads require: Larger number of spindles ( disks ) per node Not sure go with balance workloads configuration

56 Hardware Topology Hadoop uses a master / slave topology Master Nodes include: NameNode - maintains system metadata Backup NN- performs checkpoint operations and host standby ResourceManager- manages task assignment Slave Nodes include: DataNode - stores hdfs files / manages read and write requests Preferably co-located with TaskTracker NodeManager - performs map / reduce tasks

57 Sizing The Cluster Remember... Scaling is a relatively simple task Start with a moderate sized cluster Grow the cluster as requirements dictate Develop a scaling strategy As simple as scaling is adding new nodes takes time and resources Don't want to be adding new nodes each week Amount of data typically defines initial cluster size rate at which the volume of data increases Drivers for determining when to grow your cluster Storage requirements Processing requirements Memory requirements

58 Storage Reqs Drive Cluster Growth Data volume increases at a rate of 1TB / week 3TB of storage are required to store the data alone Remember block replication Consider additional overhead - typically 30% Remember files that are stored on a nodes local disk If DataNodes incorporate 4-1TB drives 1 new node per week is required 2 years of data - roughly 100TB will require 100 new nodes

59 Things Break Things are going to break This assumption is a core premise of Hadoop If a disk fails, the infrastructure must accommodate If a DataNode fails, the NameNode must manage this If a task fails, the ApplicationMaster must manage this failure Master nodes are typically a SPOF unless using a Highly Available configuration NameNode goes down, HDFS is inaccessible Use NameNode HA ResourceManager goes down, can't run any jobs Use RM HA (in development)

60 Cluster Nodes Cluster nodes should be commodity hardware Buy more nodes... Not more expensive nodes Workload patterns and cluster size drive CPU choice Small cluster - 50 nodes or less Quad core / medium clock speed is usually sufficient Large cluster Dual 8-core CPUs with a medium clock speed is sufficient Compute intensive workloads might require higher clock speeds General guideline is to buy more hardware instead of faster hardware Lots of memory - 48GB / 64GB / 128GB / 256GB Each map / reduce task consumes 1GB to 3GB of memory OS / Daemons consume memory as well

61 Cluster Storage 4 to 12 drives of 1TB / 2TB capacity - up to 24TB / node 3TB drives work Network performance penalty if a node fails 7200 rpm SATA drives are sufficient Slightly above average MTBF is advantageous JBOD configuration RAID is slow RAID is not required due to block replication More smaller disks is preferred over fewer larger disks Increased parallelism for DataNodes Slaves should never use virtual memory

62 Master Nodes Still commodity hardware, but... better Redundant everything Power supplies Dual Ethernet cards 16 to 24 CPU cores on NameNodes NameNodes and their clients are very chatty and need more cores to handle messaging traffic Medium clock speeds should be sufficient

63 Master Nodes HDFS namespace is limited to the amount of memory on the NameNode RAID and NFS storage on NameNode Typically RAID5 with hot spare Second remote directory such as NFS Quorum Journal Manager for HA

64 Network Considerations Hadoop is bandwidth intensive This can be a significant bottleneck Use dedicated switches 10Gb Ethernet is pretty good for large clusters

65 Which Operating System? Choose an OS that you are comfortable and familiar with Consider you admin resources / experience RedHat Enterprise Linux Includes support contract CentOS No support but the price is right Many other possibilities SuSE Enterprise Linux Ubuntu Fedora

66 Which Java Virtual Machine? Oracle Java is the only supported JVM Runs on OpenJDK, but use at your own risk Hadoop 1.0 requires Java JDK 1.6 or higher Hadoop 2.x requires Java JDK 1.7

67 References Give it a test drive! b11-final12.pdf basic-training-verisign