application architectures with hadoop - big data techcon sf 2014

Application Architectures with Hadoop Big Data TechCon San Francisco 2014 slideshare.com/hadooparchbook Mark Grover | @mark_grover Jonathan Seidman | @jseidman

2

About the book •  @hadooparchbook •  hadooparchitecturebook.com •  github.com/hadooparchitecturebook •  slideshare.com/hadooparchbook

©2014 Cloudera, Inc. All Rights Reserved.

3

About Us •  Mark

–  Software Engineer –  Committer on Apache Bigtop, committer and PPMC member on Apache

Sentry (incubating). –  Contributor to Hadoop, Hive, Spark, Sqoop, Flume. –  @mark_grover

•  Jonathan –  Senior Solutions Architect, Partner Enablement. –  Co-founder of Chicago Hadoop User Group and Chicago Big Data. –  [email protected] –  @jseidman


4

Case Study Clickstream Analysis

5

Analytics


6

Analytics


7

Web Logs – Combined Log Format


244.157.45.12 - - [17/Oct/2014:21:08:30 ] "GET /seatposts HTTP/1.0" 200 4463 "http://bestcyclingreviews.com/top_online_shops" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_2) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/36.0.1944.0 Safari/537.36” 244.157.45.12 - - [17/Oct/2014:21:59:59 ] "GET /Store/cart.jsp?productID=1023 HTTP/1.0" 200 3757 "http://www.casualcyclist.com" "Mozilla/5.0 (Linux; U; Android 2.3.5; en-us; HTC Vision Build/GRI40) AppleWebKit/533.1 (KHTML, like Gecko) Version/4.0 Mobile Safari/533.1”

8

Clickstream Analytics


244.157.45.12 - - [17/Oct/2014:21:08:30 ] "GET /seatposts HTTP/1.0" 200 4463 "http://bestcyclingreviews.com/top_online_shops" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_2) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/36.0.1944.0 Safari/537.36”

9

Challenges of Hadoop Implementation


10

Challenges of Hadoop Implementation


11

Hadoop Architectural Considerations •  Storage managers?

–  HDFS? HBase? •  Data storage and modeling:

–  File formats? Compression? Schema design? •  Data movement

–  How do we actually get the data into Hadoop? How do we get it out? •  Metadata

–  How do we manage data about the data? •  Data access and processing

–  How will the data be accessed once in Hadoop? How can we transform it? How do we query it?

•  Orchestration –  How do we manage the workflow for all of this?


12

Architectural Considerations Data Storage and Modeling

13

Data Modeling Considerations •  We need to consider the following in our architecture:

–  Storage layer – HDFS? HBase? Etc. –  File system schemas – how will we lay out the data? –  File formats – what storage formats to use for our data, both raw and

processed data? –  Data compression formats?


14

Architectural Considerations Data Modeling – Storage Layer

15

Data Storage Layer Choices •  Two likely choices for raw data:


16

Data Storage Layer Choices

•  Stores data directly as files •  Fast scans •  Poor random reads/writes

•  Stores data as Hfiles on HDFS

•  Slow scans •  Fast random reads/writes


17

Data Storage – Storage Manager Considerations

•  Incoming raw data: –  Processing requirements call for batch transformations across multiple

records – for example sessionization.

•  Processed data: –  Access to processed data will be via things like analytical queries – again

requiring access to multiple records.

•  We choose HDFS –  Processing needs in this case served better by fast scans.


18

Architectural Considerations Data Modeling – Raw Data Storage

19

Storage Formats – Raw Data and Processed Data


Processed Data

Raw Data

20

Data Storage – Format Considerations

Click to enter confidentiality information

Logs (plain text)

Logs (plain text)

Logs (plain text)

Logs (plain text)

Logs (plain text) Logs

(plain text)

Logs (plain text)

Logs (plain text)

21

Data Storage – Compression


snappy

Well, maybe. But not splittable.

X Splittable. Getting better…

Hmmm…. Splittable, but no...

22

Data Storage – More About Snappy – Designed at Google to provide high compression speeds with reasonable

compression. – Not the highest compression, but provides very good performance for

processing on Hadoop. – Snappy is not splittable though, which brings us to…


23

Hadoop File Types •  Formats designed specifically to store and process data on Hadoop:

–  File based – Sequence File –  Serialization formats – Thrift, Protocol Buffers, Avro –  Columnar formats – RCFile, ORC, Parquet


24

SequenceFile

• Stores records as binary key/value pairs.

• SequenceFile “blocks” can be compressed.

•  This enables splittability with non-splittable compression.


25

Avro

• Kinda SequenceFile on Steroids.

• Self-documenting – stores schema in header.

• Provides very efficient storage.

• Supports splittable compression.


26

Our Format Choices… •  Avro with Snappy

–  Snappy provides optimized compression. –  Avro provides compact storage, self-documenting files, and supports schema

evolution. –  Avro also provides better failure handling than other choices.

•  SequenceFiles would also be a good choice, and are directly supported by ingestion tools in the ecosystem. –  But only supports Java.


27

Architectural Considerations Data Modeling – Processed Data Storage

28

Storage Formats – Raw Data and Processed Data


Processed Data

Raw Data

29

Access to Processed Data


Column A Column B Column C Column D Value Value Value Value Value Value Value Value Value Value Value Value Value Value Value Value

Analytical Queries

30

Columnar Formats •  Eliminates I/O for columns that are not part of a query. •  Works well for queries that access a subset of columns. •  Often provide better compression. •  These add up to dramatically improved performance for many

queries.


1 2014-10-13

abc

2 2014-10-14

def

3 2014-10-15

ghi

1 2 3

2014-10-13

2014-10-14

2014-10-15

abc def ghi

31

Columnar Choices – RCFile •  Provides efficient processing for MapReduce applications, but

generally only used with Hive. •  Only supports Java. •  No Avro support. •  Limited compression support. •  Sub-optimal performance compared to newer columnar formats.


32

Columnar Choices – ORC •  A better RCFile. •  Supports Hive, but currently limited support for other interfaces. •  Only supports Java.


33

Columnar Choices – Parquet •  Supports multiple programming interfaces – MapReduce, Hive,

Impala, Pig. •  Multiple language support. •  Broad vendor support. •  These features make Parquet a good choice for our processed data.


34

Architectural Considerations Data Modeling – HDFS Schema Design

35

Recommended HDFS Schema Design •  How to lay out data on HDFS?


36

Recommended HDFS Schema Design /user/<username> - User specific data, jars, conf files /etl – Data in various stages of ETL workflow /tmp – temp data from tools or shared between users /data – shared data for the entire organization /app – Everything but data: UDF jars, HQL files, Oozie workflows


37

Architectural Considerations Data Modeling – Advanced HDFS Schema Design

38

What is Partitioning?

dataset col=val1/file.txt col=val2/file.txt . . . col=valn/file.txt

dataset file1.txt file2.txt . . . filen.txt

Un-partitioned HDFS directory structure

Partitioned HDFS directory structure


39

What is Partitioning?

clicks dt=2014-01-01/clicks.txt dt=2014-01-02/clicks.txt . . . dt=2014-03-31/clicks.txt

clicks clicks-2014-01-01.txt clicks-2014-01-02.txt . . . clicks-2014-03-31.txt

Un-partitioned HDFS directory structure

Partitioned HDFS directory structure


40

Partitioning •  Split the dataset into smaller consumable chunks •  Rudimentary form of “indexing” •  <data set name>/<partition_column_name=partition_column_value>/

{files}


41

Partitioning considerations •  What column to partition by?

–  HDFS is append only. –  Don’t have too many partitions (<10,000) –  Don’t have too many small files in the partitions (more than block size

generally)

•  We decided to partition by timestamp


42

Partitioning For Our Case Study •  Raw dataset:

–  /etl/BI/casualcyclist/clicks/rawlogs/year=2014/month=10/day=10!

•  Processed dataset: –  /data/bikeshop/clickstream/year=2014/month=10/day=10!


43

A Word About Partitioning Alternatives

year=2014/month=10/day=10 dt=2014-10-10


44

Architectural Considerations Data Ingestion

45

File Transfers

•  “hadoop fs –put <file>” • Reliable, but not

resilient to failure. • Other options are

mountable HDFS, for example NFSv3.


46

Streaming Ingestion •  Flume

–  Reliable, distributed, and available system for efficient collection, aggregation and movement of streaming data, e.g. logs.

•  Kafka –  Reliable and distributed publish-subscribe messaging system.


47

Flume vs. Kafka

• Purpose built for Hadoop data ingest.

• Pre-built sinks for HDFS, HBase, etc.

• Supports transformation of data in-flight.

• General pub-sub messaging framework.

• Hadoop not supported, requires 3rd-party component (Camus).

•  Just a message transport (a very fast one).


48

Flume vs. Kafka •  Bottom line:

–  Flume very well integrated with Hadoop ecosystem, well suited to ingestion of sources such as log files.

–  Kafka is a highly reliable and scalable enterprise messaging system, and great for scaling out to multiple consumers.


49

Sources Interceptors Selectors Channels Sinks

Flume Agent

Short Intro to Flume Twitter, logs, JMS,

webserver Mask, re-format,

validate… DR, critical Memory, file HDFS, HBase,

Solr

50

A Quick Introduction to Flume •  Reliable – events are stored in channel until delivered to next

stage. •  Recoverable – events can be persisted to disk and recovered in

the event of failure.

Flume Agent

Source Channel Sink Destination


51

A Quick Introduction to Flume

• Declarative – No coding required. – Configuration specifies

how components are wired together.


52

A Brief Discussion of Flume Patterns – Fan-in

•  Flume agent runs on each of our servers.

•  These agents send data to multiple agents to provide reliability.

•  Flume provides support for load balancing.


53

Sqoop Overview •  Apache project designed to ease import and export of data between

Hadoop and external data stores such as relational databases. •  Great for doing bulk imports and exports of data between HDFS,

Hive and HBase and an external data store. Not suited for ingesting event based data.


54

Ingestion Decisions •  Historical Data

–  Smaller files: file transfer –  Larger files: Flume with spooling directory source.

•  Incoming Data –  Flume with the spooling directory source.

•  Relational Data Sources – ODS, CRM, etc. –  Sqoop


55

Architectural Considerations Data Processing – Engines

56

Data flow

Raw data

Partitioned clickstream data

Other data (Financial, CRM, etc.)

Aggregated dataset #2

Aggregated dataset #1


57

Processing Engines •  MapReduce •  Abstractions – Pig, Hive, Cascading, Crunch •  Spark •  Impala

Confidentiality Information Goes Here

58

MapReduce •  Oldie but goody •  Restrictive Framework / Innovated Work Around •  Extreme Batch


59

MapReduce Basic High Level


Mapper

HDFS (Replicated)

Native File System

Block of Data

Temp Spill Data

Partitioned Sorted Data

Reducer

Reducer Local Copy

Output File

60

Abstractions •  SQL

–  Hive

•  Script/Code –  Pig: Pig Latin –  Crunch: Java/Scala –  Cascading: Java/Scala


61

Spark •  The New Kid that isn’t that New Anymore •  Easily 10x less code •  Extremely Easy and Powerful API •  Very good for machine learning •  Scala, Java, and Python •  RDDs •  DAG Engine


62

Spark - DAG


63

Impala • Real-time open source MPP style engine for Hadoop • Doesn’t build on MapReduce • Written in C++, uses LLVM for run-time code generation • Can create tables over HDFS or HBase data • Accesses Hive metastore for metadata • Access available via JDBC/ODBC


64

What processing needs to happen?


•  Sessionization •  Filtering •  Deduplication •  BI / Discovery

65

Sessionization


Website visit

Visitor 1 Session 1

Visitor 1 Session 2

Visitor 2 Session 1

> 30 minutes

66

Why sessionize?


Helps answers questions like: •  What is my website’s bounce rate?

–  i.e. how many % of visitors don’t go past the landing page?

•  Which marketing channels (e.g. organic search, display ad, etc.) are leading to most sessions? –  Which ones of those lead to most conversions (e.g. people buying things,

signing up, etc.)

•  Do attribution analysis – which channels are responsible for most conversions?

67

How to Sessionize?


1.  Given a list of clicks, determine which clicks came from the same user

2.  Given a particular user's clicks, determine if a given click is a part of a new session or a continuation of the previous session

68

#1 – Which clicks are from same user? •  We can use:

–  IP address (244.157.45.12) –  Cookies (A9A3BECE0563982D) –  IP address (244.157.45.12)and user agent string ((KHTML, like Gecko) Chrome/36.0.1944.0 Safari/537.36")


69

#1 – Which clicks are from same user? •  We can use:

–  IP address (244.157.45.12) –  Cookies (A9A3BECE0563982D) –  IP address (244.157.45.12)and user agent string ((KHTML, like Gecko) Chrome/36.0.1944.0 Safari/537.36")


70

#1 – Which clicks are from same user?



71

#2 – Which clicks part of the same session?



> 30 mins apart = different sessions

72 ©2014 Cloudera, Inc. All rights reserved.

Sessionization engine recommendation •  We have sessionization code in MR, Spark on github. The

complexity of the code varies, depends on the expertise in the organization.

•  We choose MR, since it’s fairly simple and maintainable code.

73

Filtering – filter out incomplete records


244.157.45.12 - - [17/Oct/2014:21:08:30 ] "GET /seatposts HTTP/1.0" 200 4463 "http://bestcyclingreviews.com/top_online_shops" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_2) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/36.0.1944.0 Safari/537.36” 244.157.45.12 - - [17/Oct/2014:21:59:59 ] "GET /Store/cart.jsp?productID=1023 HTTP/1.0" 200 3757 "http://www.casualcyclist.com" "Mozilla/5.0 (Linux; U…

74

Filtering – filter out records from bots/spiders



Google spider IP address


Filtering recommendation •  Bot/Spider filtering can be done easily in any of the engines •  Incomplete records are harder to filter in schema systems like

Hive, Impala, Pig, etc. •  Pretty close choice between MR, Hive and Spark •  Can be done in Flume interceptors as well •  We can simply embed this in our sessionization job

76

Deduplication – remove duplicate records


244.157.45.12 - - [17/Oct/2014:21:08:30 ] "GET /seatposts HTTP/1.0" 200 4463 "http://bestcyclingreviews.com/top_online_shops" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_2) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/36.0.1944.0 Safari/537.36” 244.157.45.12 - - [17/Oct/2014:21:08:30 ] "GET /seatposts HTTP/1.0" 200 4463 "http://bestcyclingreviews.com/top_online_shops" "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_9_2) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/36.0.1944.0 Safari/537.36”


Deduplication recommendation •  Can be done in all engines. •  We already have a Hive table with all the columns, a simple

DISTINCT query will perform deduplication •  We use Pig


BI/Discovery engine recommendation •  Main requirements for this are:

–  Low latency –  SQL interface (e.g. JDBC/ODBC) –  Users don’t know how to code

•  We chose Impala –  It’s a SQL engine –  Much faster than other engines –  Provides standard JDBC/ODBC interfaces

79

Architectural Considerations Orchestration

80

•  Data arrives through Flume •  Triggers a processing event:

–  Sessionize –  Enrich – Location, marketing channel… –  Store as Parquet

•  Each day we process events from the previous day

Orchestrating Clickstream


•  Workflow is fairly simple •  Need to trigger workflow based on data •  Be able to recover from errors •  Perhaps notify on the status •  And collect metrics for reporting

Choosing Right



Choosing…

Easier in Oozie



Choosing the right Orchestration Tool

Better in Azkaban


The best orchestration tool is the one you are an expert on

Important Decision Consideration!

86

Putting It All Together Final Architecture

87

Final Architecture – High Level Overview

Data Sources Ingestion Data Storage/

Processing

Data Reporting/Analysis


88



Processing



89

Final Architecture – Ingestion

Web App Avro Agent Web App Avro Agent




Flume Agent

Flume Agent

Flume Agent

Flume Agent

Fan-in Pattern

Multi Agents for Failover and rolling restarts

HDFS


90



Processing



91

Final Architecture – Storage and Processing

/etl/weblogs/20140331/ /etl/weblogs/20140401/ …

Data Processing

/data/marketing/clickstream/bouncerate/ /data/marketing/clickstream/attribution/ …


92



Processing



93

Final Architecture – Data Access

Hive/Impala

BI/Analytics

Tools

DWH Sqoop

Local Disk

R, etc.

DB import tool

JDBC/ODBC


Thank you

95

Contact info •  Mark Grover

–  @mark_grover –  www.linkedin.com/in/grovermark

•  Jonathan Seidman –  [email protected] –  @jseidman –  www.linkedin.com/in/jaseidman –  www.slideshare.net/jseidman

•  Slides at slideshare.net/hadooparchbook


application architectures with hadoop - big data techcon sf 2014

Technology

data access

data storage compressionclick

process data

data movement

storage formats raw

data compression formats

data modeling considerations

raw andprocessed data