iloug tech days 2016 - big data for oracle developers - towards spark, real-time and predictive...

[email protected] www.rittmanmead.com @rittmanmead

Big Data for Oracle Developers & DBAs - Towards Spark, Real-Time and Predictive Analytics

Mark Rittman, CTO, Rittman Mead IlOUG Tech Day 2016 Day 2 Keynote, 31st May 2016 @ 9.15am

[email protected] www.rittmanmead.com @rittmanmead 2

•Mark Rittman, Co-Founder of Rittman Mead‣Oracle ACE Director, specialising in Oracle BI&DW‣14 Years Experience with Oracle Technology‣Regular columnist for Oracle Magazine

•Author of two Oracle Press Oracle BI books‣Oracle Business Intelligence Developers Guide‣Oracle Exalytics Revealed‣Writer for Rittman Mead Blog :http://www.rittmanmead.com/blog

•Email : [email protected]•Twitter : @markrittman

About the Speaker

mailto:[email protected]


•Everyone’s talking about Hadoop and “Big Data”

Hadoop is the Big Hot Topic In IT / Analytics


•Gives us an ability to store more data, at more detail, for longer•Provides a cost-effective way to analyse vast amounts of data•Hadoop & NoSQL technologies can give us “schema-on-read” capabilities•There’s vast amounts of innovation in this area we can harness•And it’s very complementary to Oracle BI & DW

Why is Hadoop of Interest to Us?


Flexible Cheap Storage for Logs, Feeds + Social Data

$50k

Hadoop Node

Voice + Chat Transcripts

Call Center LogsChat Logs iBeacon Logs Website LogsCRM Data Transactions Social FeedsDemographics

Raw Data

Customer 360 Apps

Predictive Models

SQL-on-Hadoop

Business analytics

Real-time Feeds,batch and API


Incorporate Hadoop Data Reservoirs into DW Design

Virtu

aliz

atio

n &

Q

uery

Fed

erat

ion

Enterprise Performance Management

Pre-built & Ad-hoc BI Assets

Information Services

Data Ingestion

Information Interpretation

Access & Performance Layer

Foundation Data Layer

Raw Data Reservoir

Data Science

Data Engines & Poly-structured sources

Content

Docs Web & Social Media

SMS

Structured Data Sources

•Operational Data •COTS Data •Master & Ref. Data •Streaming & BAM

Immutable raw data reservoir Raw data at rest is not interpreted

Immutable modelled data. Business Process Neutral form. Abstracted from business process changes

Past, current and future interpretation of enterprise data. Structured to support agile access & navigation

Discovery Lab Sandboxes Rapid Development Sandboxes

Project based data stores to support specific discovery objectives

Project based data stored to facilitate rapid content / presentation delivery

Data Sources


•Oracle Engineered system for big data processing and analysis•Start with Oracle Big Data Appliance Starter Rack - expand up to 18 nodes per rack•Cluster racks together for horizontal scale-out using enterprise-quality infrastructure

Oracle Big Data Appliance Starter Rack + Expansion

• Cloudera CDH + Oracle software • 18 High-spec Hadoop Nodes with

InfiniBand switches for internal Hadoop traffic, optimised for network throughput

• 1 Cisco Management Switch• Single place for support for H/W + S/

W

Deployed on Oracle Big Data Appliance

Oracle Big Data Appliance Starter Rack + Expansion

• Cloudera CDH + Oracle software • 18 High-spec Hadoop Nodes with

InfiniBand switches for internal Hadoop traffic, optimised for network throughput

• 1 Cisco Management Switch• Single place for support for H/W + S/

W

Enriched Customer Profile

Modeling

Scoring

Infiniband


•Hadoop, through MapReduce, breaks processing down into simple stages‣Map : select the columns and values you’re interested in, pass through as key/value pairs‣Reduce : aggregate the results

•Most ETL jobs can be broken down into filtering, projecting and aggregating

•Hadoop then automatically runs job on cluster‣Share-nothing small chunks of work‣Run the job on the node where the data is‣Handle faults etc‣Gather the results back in

Hadoop Tenets : Simplified Distributed Processing

Mapper Filter, Project



Reducer Aggregate

Reducer Aggregate

Output One HDFS file per reducer,in a directory


•MapReduce jobs are typically written in Java, but Hive can make this simpler•Hive is a query environment over Hadoop/MapReduce to support SQL-like queries•Hive server accepts HiveQL queries via HiveODBC or HiveJDBC, automaticallycreates MapReduce jobs against data previously loaded into the Hive HDFS tables

•Approach used by ODI and OBIEEto gain access to Hadoop data

•Allows Hadoop data to be accessed just like any other data source (sort of...)

Hive as the Hadoop SQL Access Layer


•Data integration tools such as Oracle Data Integrator can load and process Hadoop data•BI tools such as Oracle Business Intelligence 12c can report on Hadoop data•Generally use MapReduce and Hive to access data‣ODBC and JDBC access to Hive tabular data‣Allows Hadoop unstructured/semi-structureddata on HDFS to be accessed like RDBMS

Hive Provides a SQL Interface for BI + ETL Tools

Access direct Hive or extract using ODI12c for structured OBIEE dashboard analysis

What pages are people visiting? Who is referring to us on Twitter? What content has the most reach?


•Most Oracle DBAs and developers know about Hadoop, but assume…

Common Developer Understanding of Hadoop Today

‣Hadoop is just for batch (because of the MapReduce JVN spin-up issue) ‣Hadoop is just for large datasets, not ad-hoc work or micro batches ‣Hadoop will always be slow because it stages everything to disk ‣All Hadoop can do is Map (select, filter) and Reduce (aggregate) ‣Hadoop == MapReduce

MapReduce is for losers

Hive is Slow

Hive is slow

too slowfor ad-hoc querying

But the future is fast

And pretty damn exciting


Hadoop 2.0 Processing Frameworks + Tools


•MapReduce’s great innovation was to break processing down into distributed jobs•Jobs that have no functional dependency on each other, only upstream tasks•Provides a framework that is infinitely scalable and very fault tolerant•Hadoop handled job scheduling and resource management‣All MapReduce code had to do was provide the “map” and “reduce” functions‣Automatic distributed processing‣Slow but extremely powerful

Hadoop 1.0 and MapReduce


•A typical Hive or Pig script compiles down into multiple MapReduce jobs•Each job stages its intermediate results to disk•Safe, but slow - write to disk, spin-up separate JVMs for each job

MapReduce - Scales By Writing Intermediate Results to Disk

SELECT LOWER(hashtags.text), COUNT(*) AS total_count FROM ( SELECT * FROM tweets WHERE regexp_extract(created_at,"(2015)*",1) = "2015" ) tweets LATERAL VIEW EXPLODE(entities.hashtags) t1 AS hashtags GROUP BY LOWER(hashtags.text) ORDER BY total_count DESC LIMIT 15

MapReduce Jobs Launched: Stage-Stage-1: Map: 1 Reduce: 1 Cumulative CPU: 5.34 sec HDFS Read: 10952994 HDFS Write: 5239 SUCCESS Stage-Stage-2: Map: 1 Reduce: 1 Cumulative CPU: 2.1 sec HDFS Read: 9983 HDFS Write: 164 SUCCESS Total MapReduce CPU Time Spent: 7 seconds 440 msec OK

1

2


•MapReduce 2 (MR2) splits the functionality of the JobTrackerby separating resource management and job scheduling/monitoring

• Introduces YARN (Yet Another Resource Manager)•Permits other processing frameworks to MR‣For example, Apache Spark

•Maintains backwards compatibility with MR1• Introduced with CDH5+

MapReduce 2 and YARN

Node Manager

Node Manager

Node Manager

Resource Manager

Client

Client


•Runs on top of YARN, provides a faster execution engine than MapReduce for Hive, Pig etc•Models processing as an entire data flow graph (DAG), rather than separate job steps‣DAG (Directed Acyclic Graph) is a new programming style for distributed systems‣Dataflow steps pass data between them as streams, rather than writing/reading from disk

•Supports in-memory computation, enables Hive on Tez (Stinger) and Pig on Tez•Favoured In-memory / Hive v2 route by Hortonworks

Apache Tez

Inpu

t Dat

a

TEZ DAGMap()

Map()

Map()

Reduce()

Out

put D

ata

Reduce()

Reduce()

Reduce()

Inpu

t Dat

a Map()

Map()

Reduce()

Reduce()


Tez Advantage - Drop-In Replacement for MR with Hive, Pig

set hive.execution.engine=mr

set hive.execution.engine=tez

4m 17s

2m 25s


•Cloudera’s answer to Hive query response time issues•MPP SQL query engine running on Hadoop, bypasses MapReduce for direct data access•Mostly in-memory, but spills to disk if required

•Uses Hive metastore to access Hive table metadata•Similar SQL dialect to Hive - not as rich though and no support for Hive SerDes, storage handlers etc

Cloudera Impala - Fast, MPP-style Access to Hadoop Data


•Beginners usually store data in HDFS using text file formats (CSV) but these have limitations•Apache AVRO often used for general-purpose processing‣Splitability, schema evolution, in-built metadata, support for block compression

•Parquet now commonly used with Impala due to column-orientated storage‣Mirrors work in RDBMS world around column-store‣Only return (project) the columns you require across a wide table

Apache Parquet - Column-Orientated Storage for Analytics


•But Parquet (and HDFS) have significant limitation for real-time analytics applications‣Append-only orientation, focus on column-store makes streaming ingestion harder

•Cloudera Kudu aims to combine best of HDFS + HBase‣Real-time analytics-optimised ‣Supports updates to data‣Fast ingestion of data‣Accessed using SQL-style tablesand get/put/update/delete API

Cloudera Kudu - Combining Best of HBase and Column-Store


•Part of Oracle Big Data 4.0 (BDA-only)‣Also requires Oracle Database 12c, Oracle Exadata Database Machine

•Extends Oracle Data Dictionary to cover Hive•Extends Oracle SQL and SmartScan to Hadoop•Extends Oracle Security Model over Hadoop‣Fine-grained access control‣Data redaction, data masking‣Uses fast c-based readers where possible(vs. Hive MapReduce generation)‣Map Hadoop parallelism to Oracle PQ‣Big Data SQL engine works on top of YARN‣Like Spark, Tez, MR2

Oracle Big Data SQL

Exadata Storage Servers

HadoopCluster

Exadata DatabaseServer

Oracle Big Data SQL

SQL Queries

SmartScan SmartScan

hold on…

“where we’re goingwe don’t need roads”


•Apache Drill is another SQL-on-Hadoop project that focus on schema-free data discovery• Inspired by Google Dremel, innovation is querying raw data with schema optional

•Automatically infers and detects schema from semi-structured datasets and NoSQL DBs•Join across different silos of data e.g. JSON records, Hive tables and HBase database•Aimed at different use-cases than Hive - low-latency queries, discovery (think Endeca vs OBIEE)

Introducing Apache Drill - “We Don’t Need No Roads”


•Most modern datasource formats embed their schema in the data (“schema-on-read”)•Apache Drill makes these as easy to join to traditional datasets as “point me at the data”

•Cuts out unnecessary work in defining Hive schemas for data that’s self-describing•Supports joining across files,databases, NoSQL etc

Self-Describing Data - Parquet, AVRO, JSON etc


•Files can exist either on the local filesystem, or on HDFS•Connection to directory or file defined in storage configuration

•Can work with CSV, TXT, TSV etc•First row of file can provide schema (column names)

Apache Drill and Text Files

SELECT * FROM dfs.`/tmp/csv_with_header.csv2`;+-------+------+------+------+| name | num1 | num2 | num3 |+-------+------+------+------+| hello | 1 | 2 | 3 || hello | 1 | 2 | 3 || hello | 1 | 2 | 3 || hello | 1 | 2 | 3 || hello | 1 | 2 | 3 || hello | 1 | 2 | 3 || hello | 1 | 2 | 3 |+-------+------+------+------+7 rows selected (0.12 seconds)

SELECT * FROM dfs.`/tmp/csv_no_header.csv`;+------------------------+| columns |+------------------------+| ["hello","1","2","3"] || ["hello","1","2","3"] || ["hello","1","2","3"] || ["hello","1","2","3"] || ["hello","1","2","3"] || ["hello","1","2","3"] || ["hello","1","2","3"] |+------------------------+7 rows selected (0.112 seconds)


•JSON (Javascript Object Notation) documents are often used for data interchange•Exports from Twitter and other consumer services•Web service responses and other B2B interfaces

•A more lightweight form of XML that is “self-describing”

•Handles evolving schemas, and optional attributes•Drill treats each document as a row, and has features to•Flatten nested data (extract elements from arrays)•Generate key/value pairs for loosely structured data

Apache Drill and JSON Documents

use dfs.iot;show files;select in_reply_to_user_id, text from `all_tweets.json`limit 5;+---------------------+------+| in_reply_to_user_id | text |+---------------------+------+| null | BI Forum 2013 in Brighton has now sold-out || null | "Football has become a numbers game || null | Just bought Lyndsay Wise’s Book || null | An Oracle BI "Blast from the Past" || 14716125 | Dilbert on Agile Programming |+---------------------+------+5 rows selected (0.229 seconds)

select name, flatten(fillings) as f from dfs.users.`/donuts.json` where f.cal < 300;


•Drill can connect to Hive to make use of metastore (incl. multiple Hive metastores)•NoSQL databases (HBase etc)•Parquet files (native storage format - columnar + self describing)

Apache Drill and Hive, HBase, Parquet Sources etc

USE hbase;SELECT * FROM students;

+-------------+-----------------------+-----------------------------------------------------+| row_key | account | address |+-------------+-----------------------+------------------------------------------------------+| [B@e6d9eb7 | {"name":"QWxpY2U="} | {"state":"Q0E=","street":"MTIzIEJhbGxtZXIgQXY="} || [B@2823a2b4 | {"name":"Qm9i"} | {"state":"Q0E=","street":"MSBJbmZpbml0ZSBMb29w"} || [B@3b8eec02 | {"name":"RnJhbms="} | {"state":"Q0E=","street":"NDM1IFdhbGtlciBDdA=="} || [B@242895da | {"name":"TWFyeQ=="} | {"state":"Q0E=","street":"NTYgU291dGhlcm4gUGt3eQ=="} |+-------------+-----------------------+----------------------------------------------------------------------+

SELECT firstname,lastname FROM hiveremote.`customers` limit 10;`

+------------+------------+| firstname | lastname |+------------+------------+| Essie | Vaill || Cruz | Roudabush || Billie | Tinnes || Zackary | Mockus || Rosemarie | Fifield || Bernard | Laboy || Marianne | Earman |+------------+------------+

SELECT * FROM dfs.`iot_demo/geodata/region.parquet`;

+--------------+--------------+-----------------------+| R_REGIONKEY | R_NAME | R_COMMENT |+--------------+--------------+-----------------------+| 0 | AFRICA | lar deposits. blithe || 1 | AMERICA | hs use ironic, even || 2 | ASIA | ges. thinly even pin || 3 | EUROPE | ly final courts cajo || 4 | MIDDLE EAST | uickly special accou |+--------------+--------------+-----------------------+


•Drill developed for real-time, ad-hoc data exploration with schema discovery on-the-fly • Individual analysts exploring new datasets, leveraging corporate metadata/data to help•Hive is more about large-scale, centrally curated set-based big data access

•Drill models conceptually as JSON, vs. Hive’s tabular approach•Drill introspects schema from whatever it connects to, vs. formal modeling in Hive

Apache Drill vs. Apache Hive

Interactive Queries (Data Discovery, Tableau/VA)

Reporting Queries(Canned Reports, OBIEE)

ETL (ODI, Scripting, Informatica)

Apache Drill Apache Hive

Interactive Queries 100ms - 3mins

Reporting Queries 3mins - 20mins

ETL & Batch Queries 20mins - hours

but…

what’s all this about “Spark”?

No not that “Sparc"


•Another DAG execution engine running on YARN•More mature than TEZ, with richer API and more vendor support•Uses concept of an RDD (Resilient Distributed Dataset)‣RDDs like tables or Pig relations, but can be cached in-memory‣Great for in-memory transformations, or iterative/cyclic processes

•Spark jobs comprise of a DAG of tasks operating on RDDs•Access through Scala, Python or Java APIs•Related projects include‣Spark SQL‣Spark Streaming

Apache Spark


•Native support for multiple languages with identical APIs‣Python - prototyping, data wrangling‣Scala - functional programming features‣Java - lower-level, application integration

•Use of closures, iterations, and other common language constructs to minimize code

• Integrated support for distributed +functional programming

•Unified API for batch and streaming

Rich Developer Support + Wide Developer Ecosystem

scala> val logfile = sc.textFile("logs/access_log") 14/05/12 21:18:59 INFO MemoryStore: ensureFreeSpace(77353) called with curMem=234759, maxMem=309225062 14/05/12 21:18:59 INFO MemoryStore: Block broadcast_2 stored as values to memory (estimated size 75.5 KB, free 294.6 MB) logfile: org.apache.spark.rdd.RDD[String] = MappedRDD[31] at textFile at <console>:15 scala> logfile.count() 14/05/12 21:19:06 INFO FileInputFormat: Total input paths to process : 1 14/05/12 21:19:06 INFO SparkContext: Starting job: count at <console>:1 ... 14/05/12 21:19:06 INFO SparkContext: Job finished: count at <console>:18, took 0.192536694 s res7: Long = 154563

scala> val logfile = sc.textFile("logs/access_log").cache scala> val biapps11g = logfile.filter(line => line.contains("/biapps11g/")) biapps11g: org.apache.spark.rdd.RDD[String] = FilteredRDD[34] at filter at <console>:17 scala> biapps11g.count() ... 14/05/12 21:28:28 INFO SparkContext: Job finished: count at <console>:20, took 0.387960876 s res9: Long = 403


•Spark SQL, and Data Frames, allow RDDs in Spark to be processed using SQL queries•Bring in and federate additional data from JDBC sources•Load, read and save data in Hive, Parquet and other structured tabular formats

Spark SQL - Adding SQL Processing to Apache Spark

val accessLogsFilteredDF = accessLogs .filter( r => ! r.agent.matches(".*(spider|robot|bot|slurp).*")) .filter( r => ! r.endpoint.matches(".*(wp-content|wp-admin).*")).toDF() .registerTempTable("accessLogsFiltered") val topTenPostsLast24Hour = sqlContext.sql("SELECT p.POST_TITLE, p.POST_AUTHOR, COUNT(*) as total FROM accessLogsFiltered a JOIN posts p ON a.endpoint = p.POST_SLUG GROUP BY p.POST_TITLE, p.POST_AUTHOR ORDER BY total DESC LIMIT 10 ") // Persist top ten table for this window to HDFS as parquet file topTenPostsLast24Hour.save("/user/oracle/rm_logs_batch_output/topTenPostsLast24Hour.parquet" , "parquet", SaveMode.Overwrite)

Hadoop is insecureand has fragmented security

…doesn’t it?

but …


Consistent Security and Audit Now Emerging on Platform


•Clusters by default are unsecured (vunerable to account spoofing) & need Kerberos enabled•Data access controlled by POSIX-style permissions on HDFS files•Hive and Impala can Apache Sentry RBAC‣Result is data duplication and complexity‣No consistent API or abstracted security model

Hadoop Security Initially Was a Mess

/user/mrittman/scratchpad /user/ryeardley/scratchpad /user/mpatel/scratchpad /user/mrittman/scratchpad /user/mrittman/scratchpad /data/rm_website_analysis/logfiles/incoming /data/rm_website_analysis/logfiles/archive /data/rm_website_analysis/tweets/incoming /data/rm_website_analysis/tweets/archive


•Use standard Oracle Security over Hadoop & NoSQL‣Grant & Revoke Privileges‣Redact Data‣Apply Virtual Private Database‣Provides Fine-grain Access Control

•Great solution to extend existing Oraclesecurity model over Hadoop datasets

Oracle Big Data SQL : Extend Oracle Security to Hadoop

Redacteddatasubset

SQLJSON

CustomerdatainOracleDB

DBMS_REDACT.ADD_POLICY( object_schema => 'txadp_hive_01', object_name => 'customer_address_ext', column_name => 'ca_street_name', policy_name => 'customer_address_redaction', function_type => DBMS_REDACT.RANDOM, expression => 'SYS_CONTEXT(''SYS_SESSION_ROLES'', ''REDACTION_TESTER'')=''TRUE''' );


•Provides a higher level, logical abstraction for data (ie Tables or Views) ‣Can be used with Spark & Spark SQL, with Predicate pushdown, projection

•Returns schemed objects (instead of paths and bytes) in similar way to HCatalog•Unified data access path allows platform-wide performance improvements•Secure service that does not execute arbitrary user code‣Central location for all authorization checks using Sentry metadata.

Cloudera RecordService

any predictive modellinghas to be done outside Hadoop, in R

…doesn’t it?

but …


•Part of Spark, extends Scala, Java & Python API• Integrated workflow including ML pipelines•Currently supports following algorithms:‣Binary classification‣Regression‣Clustering‣Collaborative filtering‣Dimensionality Reduction

Spark MLLib : Adding Machine Learning Capabilities to Spark

// Compute raw scores on the test set. val scoreAndLabels = test.map { point => val score = model.predict(point.features) (score, point.label) }

// Get evaluation metrics. val metrics = new BinaryClassificationMetrics(scoreAndLabels) val auROC = metrics.areaUnderROC()

println("Area under ROC = " + auROC)

// Save and load model model.save(sc, "myModelPath") val sameModel = SVMModel.load(sc, "myModelPath")


•Data enrichment tool aimed at domain experts, not programmers •Uses machine-learning to automate data classification + profiling steps

•Automatically highlight sensitive data,and offer to redact or obfuscate

•Dramatically reduce the time requiredto onboard new data sources

•Hosted in Oracle Cloud for zero-install‣File upload and download from browser‣Automate for production data loads

Raw Data

Data stored in the original format (usually files) such as SS7, ASN.

1, JSON etc.

Mapped Data

Data sets produced by mapping and

transforming raw data

Voice + Chat Transcripts

Example Usage : Oracle Big Data Preparation Cloud Service


Identifying Schemas in Semi-/Unstructured Data


Use of Machine Learning to Identify Data Patterns

•Automatically profile, parse and classify incoming datasets using Spark MLLib Word2Vec •Spot and obfuscate sensitive data automatically, automatically suggest column names


•Hadoop is evolving‣Hadoop 2.0 breaks the dependency on MapReduce‣Spark, Tez etc allow us to create execution plans that run in-memory, faster than before‣New streaming models allow us to process data via sockets, micro batches or continuously

•And Oracle developers can make use of these new capabilities‣Oracle Big Data SQL can access Hadoop data loaded in real-time‣OBIEE, particularly in 11.1.1.9, can access Impala‣ODI is likely to support Hive on Tez and Hive on Spark shortly, and will have support for Spark in the future

Summary


Big Data for Oracle Devs - Towards Spark, Real-Time and Predictive Analytics

Mark Rittman, CTO, Rittman Mead Riga Dev Day 2016, Riga, March 2016

iloug tech days 2016 - big data for oracle developers - towards spark, real-time and predictive...

Data & Analytics