Share
Embed
Open Source Big Data in the Cloud (HDFS, Map Reduce, Hive, Spark, Kafka)
Getting started with Hadoop with HDFS, Map/Reduce, Hive, Spark and Kafka using Oracle Big Data Cloud Service Compute Edition
Frank Munz
March 21, 2018
Transcript
52 slides
-
Edelweiss Kammermann Frank Munz Open Source Big Data in OPC
Analytics & Big Data Summit 2018 -
© IT Convergence 2016. All rights reserved. About Me à
Computer Engineer, BI and Data Integration Specialist à Over 20 years of Consulting and Project Management experience in Oracle tech. à Co-founder and Vice President of Uruguayan Oracle User Group (UYOUG) à Vice President of LAOUC (Latin America Oracle User Community) à BI Manager at ITConvergence à Writer and frequent speaker at tech conferences à Oracle ACE Director à Oracle Big Data Implementation Specialist -
6 Dr. Frank Munz •Founded munz & more in 2007
•18 years Oracle Middleware, Cloud, and Distributed Computing •Consulting and High-End Training •Wrote two Oracle WLS and one Cloud book -
© IT Convergence 2016. All rights reserved. What is Big
Data? ü Volume: High amount of data ü Variety: Different data types formats. Unstructured/semi-structured data ü Velocity: Speed which data is created and/or consumed ü Veracity: Quality of data. Accuracy ü Value: Data has intrinsic value—but it must be discovered -
© IT Convergence 2016. All rights reserved. What is Oracle
Big Data Cloud Compute Edition? à Big Data Platform that integrates Oracle Big Data solution with Open Source tools à Fully Elastic à Integrated with Other Paas Services as Database Cloud Service, MySQL Cloud Service, Event Hub Cloud Service à Access, Data and Network Security à REST access to all the funcitonality -
© IT Convergence 2016. All rights reserved. Big Data Cloud
Service – Compute Edition (BDCS-CE) -
© IT Convergence 2016. All rights reserved. BDCS-CE Notebook: Interactive
Analysis à Apache Zeppelin Notebook (version0.7) to interactively work with data -
© IT Convergence 2016. All rights reserved. What is Hadoop?
à An open source software platform for distributed storage and processing à Manage huge volumes of unstructured data à Parallel processing of large data set à Highly scalable à Fault-tolerant à Two main components: à HDFS: Hadoop Distributed File System for storing information à MapReduce: programming framework that process information -
© IT Convergence 2016. All rights reserved. HFDS Architecture (Simplified)
Client NameNode DataNodes Manages metadata and access control Has the info of where the data is (which DataNodes contains the blocks of each file) Keeps this info in memory. Store and retrieves data (blocks) by client request. Requests processes as read or write data -
© IT Convergence 2016. All rights reserved. HFDS: Writing Data
Client NameNode DataNodes 1 2 Divide the file into fixed size blocks (usually 64 or 128MB) For each block: Ask Namenode in which DataNodes can write, Specifying block size and replication factor For each block: Provide DataNodes addresses, sorted in increasing distance 3 -
© IT Convergence 2016. All rights reserved. HFDS: Writing Data
Client NameNode DataNodes 1 2 Sends the data of the block and the list of nodes to the first DataNode 3 4 5 Sends the data to the following DataNode Replication Pipeline 6 Each DataNode sends Done to NameNode once the block data is written to hard disk -
© IT Convergence 2016. All rights reserved. HFDS: Reading Data
Client NameNode DataNode 1 Send list of blocks of the file. List of DataNodes for each block 2 4 Send data for required block Ask NameNode for a specific file 3 Download data from the nearest DataNode (send block number) -
© IT Convergence 2016. All rights reserved. HFDS: High Availability
ü Secondary NameNode (active-standby configuration) ü Namenodes use shared storage ü Datanodes send block reports to both namenodes Shared Storage Passive NameNode Active NameNode -
© IT Convergence 2016. All rights reserved. HFDS: Command Examples
ü hadoop fs –ls ü hadoop fs -put <local_path> <hdfs_path> ü hadoop fs -get <hdfs_path> <local_path> ü hadoop fs -cat <hdfs_path> ü hadoop fs -rmr <hdfs_path> ü hadoop fs –copyFromLocal <local_path> <hdfs_path> -
© IT Convergence 2016. All rights reserved. Hadoop Components: MapReduce
à Retrieves data from HDFS à A MapReduce program is composed by à Map() method: performs filtering and sorting of the <key, value> inputs à Reduce() method: summarize the <key,value> pairs provided by the Mappers à Code can be written in many languages (Perl, Python, Java etc) -
HDFS & M/R • HDFS enabled big data – Democratization
• Could you operate Hadoop on-premises? • Hadoop on VirtualBox Oracle BD Light is an easy start – Localhost only munz & more #25 -
HDFS & M/R • S3 is very popular today –
Blob (key, value) storage cloud service – Not a file system, but used instead of HDFS • M/R is powerful, but low level • Google: “We do not use M/R internally at Google anymore” munz & more #26 -
© IT Convergence 2016. All rights reserved. What is Hive?
à An open source data warehouse software on top of Apache Hadoop à Analyze and query data stored in HDFS à Structure the data into tables à Tools for simple ETL à SQL- like queries (HiveQL) à Procedural language with HPL-SQL à Metadata storage in a RDBMS -
© IT Convergence 2016. All rights reserved. Hive: Code Example
à SELECT [ALL | DISTINCT] select_expr, select_expr, ... à FROM table_reference à [WHERE where_condition] à [GROUP BY col_list] à [HAVING having_condition] à [CLUSTER BY col_list | [DISTRIBUTE BY col_list] [SORT BY col_list]] à [LIMIT number]; -
Hive • Yes, SQL again! – People already know it
– Hive is basis for integration with other tools munz & more #31 -
Revisited: Map Reduce I/O munz & more #34 Source: Hadoop
Application Architecture Book -
Spark • InMemory: Orders of magnitude(s) faster than M/R •
Higher level Scala, Java, R or Python API • Standalone, in Hadoop, or Mesos • Principle: Run an operation on all data -> ”Spark is the new MapReduce” • See also: Apache Storm, etc • Uses RDDs, or Dataframes, or Datasets munz & more #35 https://stackoverflow.com/questions/31508083/difference-between- dataframe-in-spark-2-0-i-e-datasetrow-and-rdd-in-spark https://www.usenix.org/system/files/conference/nsdi12/nsdi12-final138.pdf -
RDDs Resilient Distributed Datasets Where do they come from? Collection
of data grouped into named columns. Supports text, JSON, Apache Parquet, sequence. Read in HDFS, Local FS, S3, Hbase Parallelize existing Collection Transform other RDD -> RDDs are immutable -
Lazy Evaluation munz & more #37 Nothing is executed Execution
Transformations: map(), flatMap(), reduceByKey(), groupByKey() Actions: collect(), count(), first(), takeOrdered(), saveAsTextFile(), … http://spark.apache.org/docs/2.1.1/programming-guide.html -
Word Count and Histogram munz & more #43 res =
t.flatMap(lambda line: line.split(" ")) .map(lambda word: (word, 1)) .reduceByKey(lambda a, b: a + b) res.takeOrdered(5, key = lambda x: -x[1]) -
Kafka Partitioned, replicated commit log munz & more #47 0
1 2 3 4 … n Immutable log: Messages with offset Producer Consumer A Consumer B https://www.quora.com/Kafka-writes-every-message-to-broker-disk-Still-performance-wise-it- is-better-than-some-of-the-in-memory-message-storing-message-queues-Why-is-that -
Broker1 Broker2 Broker3 Topic A (1) Topic A (2) Topic
A (3) Partition / Leader Repl A (1) Repl A (2) Repl A (3) Producer Replication / Follower Zoo- keeper Zoo- keeper Zoo- keeper State / HA -
https://www.confluent.io/blog/publishing-apache-kafka-new-york-times/ - 1 topic - 1 partition - Contains every
article published since 1851 - Multiple producers / consumers Example for Stream / Table Duality -
Kafka Clients SDKs Connect Streams - OOTB: Java, Scala -
Confluent: Python, C, C++ Confluent: - HDFS sink, - JDBC source, - S3 sink - Elastic search sink - Plugin .jar file - JDBC: Change data capture (CDC) - Real-time data ingestion - Microservices - KSQL: SQL streaming engine for streaming ETL, anomaly detection, monitoring - .jar file runs anywhere High / low level Kafka API Configuration only Integrate external Systems Data in Motion Stream / Table duality REST - Language agnostic - Easy for mobile apps - Easy to tunnel through FW etc. Lightweight -
Oracle Event Hub Cloud Service • PaaS: Managed Kafka 0.10.2
• Two deployment modes – Basic (Broker and ZK on 1 node) – Recommended (distributed) • REST Proxy – Separate sever(s) running REST Proxy munz & more #51 -
Ports You must open ports to allow access for external
clients • Kafka Broker (from OPC connect string) • Zookeeper with port 2181 munz & more #54 -
Interesting to Know • Event Hub topics are prefixed with
ID domain • With Kafka CLI topics with ID Domain can be created • Topics without ID domain are not shown in OPC console 57 -
TL;DR #bigData #openSource #OPC OpenSource: entry point to Oracle Big
Data CS world and vice versa / Low(er) setup times / Check for resource usage & limits in Big Data OPC / BDCS-CE: managed Hadoop, Hive, Spark + Event hub: Kafka / Attend a hands-on Oracle Cloud workshop! / Next level: Oracle Big Data tools @EdelweissK @FrankMunz -
3 Membership Tiers • Oracle ACE Director • Oracle ACE
• Oracle ACE Associate bit.ly/OracleACEProgram 500+ Technical Experts Helping Peers Globally Connect: Nominate yourself or someone you know: acenomination.oracle.com @oracleace Facebook.com/oracleaces [email protected]