lecture @dhbw: data warehouse 02 dwh and big data...

A company of Daimler AG

LECTURE @DHBW: DATA WAREHOUSE

02 DWH AND BIG DATA ARCHITECTUREANDREAS BUCKENHOFER, DAIMLER TSS

ABOUT ME

Andreas BuckenhoferSenior DB Professional

Since 2009 at Daimler TSS Department: Machine Learning SolutionsBusiness Unit: Analytics

DHBWDOAG

xing

Contact/Connect

vcard

• Oracle ACE Associate• DOAG responsible for InMemory DB• Lecturer at DHBW• Certified Data Vault Practitioner 2.0• Certified Oracle Professional• Certified IBM Big Data Architect

• Over 20 years experience with database technologies

• Over 20 years experience with Data Warehousing

• International project experience

mailto:[email protected]

https://twitter.com/ABuckenhofer

https://de.linkedin.com/in/buckenhofer

http://wwwlehre.dhbw-stuttgart.de/~buckenhofer/



https://www.doag.org/de/themen/datenbank/in-memory/


https://www.xing.com/profile/Andreas_Buckenhofer2

https://apex.oracle.com/pls/apex/f?p=19297:3:16881272565360:

https://www.doag.org/de/themen/datenbank/in-memory/


• Describe different DWH architectures

• Explain Big Data architectures

• Understand Data Lakes and other buzzwords

DWH LECTURE - LEARNING TARGETS

Data Warehouse / DHBWDaimler TSS 3

• Specific implementation can follow an architecture

• Architecture describes an ideal type. Therefore an implementation may not use all components or can combine components

• Better understanding, overview and complexity reduction by decomposing a DWH into its components• Can be used in many projects: repeatable, standardizable

• Map DWH tools into the different components and compare functionality

• Functional oriented as it describes data and control flow

PURPOSE: WHY ARE DWH ARCHITECTURES USEFUL?


LOGICAL STANDARD DATA WAREHOUSE ARCHITECTURE


Data Warehouse

FrontendBackend

External data sources

Internal data sources

Staging Layer(Input Layer)

OLTP

OLTP

Core Warehouse

Layer(Storage

Layer)

Mart Layer(Output Layer)

(Reporting Layer)

Integration Layer

(Cleansing Layer)

Aggregation Layer

Metadata Management

Security

DWH Manager incl. Monitor

• Providing internal and external data out of the source systems

• Enabling data through Push (source is generating extracts) or Pull (BI Data Backend is requesting or directly accessing data)

• Example for Push practice (deliver csv or text data through file interface; Change Data Capture (CDC))

• Example for Pull practice (direct access to the source system via ODBC, JDBC, API and so on)

DATA SOURCES


• “Landing Zone” for data coming into a DWH

• Purpose is to increase speed into DWH and decouple source and target system (repeating extraction run, additional delivery)

• Granular data (no pre-aggregation or filtering in the Data Source Layer, i.e. the source system)

• Usually not persistent, therefore regular housekeeping is necessary (for instance delete data in this layer that is few days/weeks old or – more common - if a correct upload to Core Warehouse Layer is ensured)

• Tables have no referential integrity constraints, columns often varchar

STAGING LAYER


• Business Rules, harmonization and standardization of data

• Classical Layer for transformations: ETL = Extract – TRANSFORM – Load

• Fixing data quality issues

• Usually not persistent, therefore regular housekeeping is necessary (for instance after a few days or weeks or at the latest once a correct upload to Core Warehouse Layer is ensured)

• The component is often not required or often not a physical part of a DB

INTEGRATION LAYER


• Data storage in an integrated, consolidated, consistent and non-redundant (normalized) data model

• Contains enterprise-wide data organized around multiple subject-areas

• Application / Reporting neutral data storage on the most detailed level of granularity (incl. historic data)

• Size of database can be several TB and can grow rapidly due to data historization

• Write-optimised layer

CORE WAREHOUSE LAYER


• Preparing data for the Data Mart Layer to the required granularity

• E.g. Aggregating daily data to monthly summaries

• E.g. Filtering data (just last 2 years or just data for a specific region)

• Harmonize computation of key performance indicators (measures) and additional Business Rules

• The component is often not required or often not a physical part of a DB

AGGREGATION LAYER


• Read-optimised layer: Data is stored in a denormalized data model for performance reasons and better end user usability/understanding

• The Data Mart Layer is providing typically aggregated data or data with less history (e.g. latest years only) in a denormalized data model

• Created through filtering or aggregating the Core Warehouse Layer

• One Mart ideally represents one subject area

• Technically the Data Mart Layer can also be a part of an Analytical Frontend product (such as Qlik, Tableau, or IBM Cognos TM1) and need not to be stored in a relational database

DATA MART LAYER


• Metadata Management

• Not just “Data about Data”, separate lecture

• Security

• Not all users are allowed to see all data

• Data security classification (e.g. restricted, confidential, secret)

• DWH Manager incl. Monitor

• DWH Manager initiates, controls, and checks job execution

• Monitor identifies changes/new data from source systems, separate lecture

METADATA MANAGEMENT, SECURITY, MONITOR


The article

http://www.kimballgroup.com/2004/03/differences-of-opinion/

compares THE two classic DWH architectures.

Read the paper and complete the table / questions on the next slide.

(Caution: The paper is biased / favors one approach; you may want to read other/more papers for a neutral view.)

EXERCISE: CLASSICAL DWH ARCHITECTURES





How are the approaches called?

Who “invented” the approach?

How many layers are used and how are the layers called?

Which data modeling approaches are used in which layer?

In which layer are atomic detail data stored?

In which layer are aggregated / summary data stored?

List at least 2 advantages

List at least 2 disadvantages



How are the approaches called?

Kimball Bus Architecture Corporate Information Factory

Who “invented” the approach?

• Ralph Kimball • Bill Inmon

How many layers are used and how are the layers called?

• Data Staging• Dimensional Data Warehouse

• Data Acquisition• Normalized Data Warehouse• Data Delivery / Dimensional Mart

Which data modeling approaches are used in which layer?

• Data Staging: variable, corresponds to source system

• Dimensional Data Warehouse:Dimensional Model

• Data Acquisition: variable, corresponds to source system

• Normalized Data Warehouse: 3NF• Data Delivery: Dimensional Model

In which layer are atomic detail data stored?

• Dimensional Data Warehouse • Normalized Data Warehouse

In which layer are aggregated / summary data stored?

• Dimensional Data Warehouse • Data Delivery / Dimensional Mart



Kimball Bus Architecture Corporate Information Factory

Advantages • Two layers only mean faster development and less work

• Rather simple approach to make data fast and easily accessible

• Lower startup costs (but higher subsequent development costs)

• Separation of concerns: long-term enterprise data storage separated from data presentation

• Changes in requirements and scope are easier to manage

• Lower subsequent development costs (but higher startup costs)

Disadvantages • If table structures change (instable source systems), high effort to implement the changes and reload data, especially conformed dimensions (“Dimensionitis” desease)

• Non-metric data not optimal for dimensional model

• Dimensional model (esp. Star Schema) contains data redundancy

• Data model transformations from 3NF to Dimensional model required

• More complex as two different data models are required

• Larger team(s) of specialists required

• Kimball Bus Architecture (Central data warehouse based on data marts)

• Inmon Corporate Information Factory

• Data Vault 2.0 Architecture (Dan Linstedt)

• DW 2.0: The Architecture for the Next Generation of Data Warehousing

• Operational Data Store (ODS)

OTHER DWH ARCHITECTURES


KIMBALL BUS ARCHITECTURE (CENTRAL DATA WAREHOUSE BASED ON DATA MARTS)


Source: http://www.kimballgroup.com/2004/03/differences-of-opinion/




Data Warehouse

FrontendBackend




OLTP

OLTP

Core Warehouse Layer= Mart Layer

Data Mart 1

Data Mart 2Data Mart 3

Metadata Management

Security


More Business-process oriented

than subject-oriented,

integrated, time-variant,non-volatile

• Bottom-up approach

• Dimensional model with denormalized data

• Sum of the data marts constitute the Enterprise DWH

• Enterprise Service Bus / conformed dimensions for integration purposes• (don’t confuse with ESB as middleware/communication system between applications)

• Kimball describes that agreeing on conformed dimensions is a hard job and it’s expected that the team will get stuck from time to time trying to align the incompatible original vocabularies of different groups

• Data marts need to be redesigned if incompatibilities exist



Co

re W

are

ho

use

La

yer

DATA INTEGRATION WITH AND WITHOUT COREWAREHOUSE LAYER


INMON CORPORATE INFORMATION FACTORY


Source: http://www.kimballgroup.com/2004/03/differences-of-opinion/

Not relevant for exam



Data Warehouse

FrontendBackend




OLTP

OLTP

Core Warehouse

Layer(Storage

Layer)


(Reporting Layer)

Metadata Management

Security


subject-oriented,

integrated, time-

variant,non-

volatile


• Top-down approach

• (Normalized) Core Warehouse is essential for subject-oriented, integrated, time-variant and nonvolatile data storage

• Create (departmental) Data Marts as subsets of Core Enterprise DWH as needed

• Data Marts can be designed with Dimensional model

• The logical standard architecture is more general compared to CIF, but was mainly influenced by CIF




DATA VAULT 2.0 ARCHITECTURE – TODAY’S WORLD (DANLINSTEDT)

Data Warehouse / DHBWDaimler TSS

DATA VAULT 2.0 ARCHITECTURE (DAN LINSTEDT)


Michael Olschimke, Dan Linstedt: Building a Scalable Data Warehouse with Data Vault 2.0, Morgan Kaufmann, 2015, Chapter 2.2



Data Warehouse

FrontendBackend




OLTP

OLTP

Raw Data Vault


(Reporting Layer)

Business Data Vault

Metadata Management

Security


Hard Rules only

Soft Rules

• Core Warehouse Layer is modeled with Data Vault and integrates data by BK (business key) “only” (Data Vault modeling is a separate lecture)

• Business rules (Soft Rules) are applied from Raw Data Vault Layer to Mart Layer and not earlier

• Alternatively from Raw Data Vault to additional layer called Business Data Vault

• Hard Rules don’t change data

• Data is fully auditable

• Real-time capable architecture

• Architecture got very popular recently; also applicable to BigData, NoSQL



• In the classical DWHs, the Core Warehouse Layer is regarded as “single version of the truth”

• Integrates + cleanses data from different sources and eliminates contradiction

• Produces consistent results/reports across Data Marts

• But: cleansing is (still) objective, Enterprises change regularly, paradigm does not scale as more and more systems exist

• Data in Raw Data Vault Layer is regarded as “Single version of the facts”

• 100% of data is loaded 100% of time

• Data is not cleansed and bad data is not removed in the Core Layer (Raw Vault)



• Data Vault is optimized for the following requirements:

• Flexibility

• Agility

• Data historization

• Data integration

• Auditability

• Bill Inmon wrote in 2008: “Data Vault is the optimal approach for modeling the EDW in the DW2.0 framework.” (DW2.0)



DW 2.0: THE ARCHITECTURE FOR THE NEXT GENERATION OF DATA WAREHOUSING


Source: W.H. Inmon, Dan Linstedt: Data Architecture: A Primer for the Data Scientist, Morgan Kaufmann, 2014, chapter 3.1

Operational applicationdata model

Integrated corporatedata model

Integrated corporatedata model

Archivaldata model

Dat

a Li

fecy

cle


Main characteristics:

• Structured and “unstructured” data, not just metrics

• Life Cycle of data with different storage areas

• Hot data: High speed, expensive storage (RAM, SSDs) for most recent data

• …

• Cold data: Low speed, inexpensive storage (e.g. hard disks) for old data; archival data model with high compression

• Metadata is an integral part of the DWH and not an afterthought

DW 2.0: THE ARCHITECTURE FOR THE NEXT GENERATION OF DATA WAREHOUSING



OPERATIONAL DATA STORE (ODS)


Data Warehouse

FrontendBackend




OLTP

OLTP

Core Warehouse

Layer(Storage

Layer)


(Reporting Layer)

Metadata Management

Security


subject-oriented,

integrated, time-

variant,non-

volatile

Operational Data Store


• ODS: Real-time/Right-time layer

• Replication techniques used to transport data from source database to ODS layer with minimal impact on source system

• Data in the ODS has no history and is stored without any cleansing and without any integration (1:1 copy from single source)

• DWH performance not optimal as data model is suited for OLTP and not for reporting requirements

• ODS normally additionally to Staging / Core DWH / Mart Layer but can exist alone without other layers

OPERATIONAL DATA STORE (ODS)



EXAMPLE DWH FOR STATE OF CONSTRUCTION DOCU



ARCHITECTURE FROM AN ACTUAL PROJECT IN THE AUTOMOTIVE INDUSTRY


ETL Engine

Fron

tend

StandardReports

AdHocReportsLogs

TSM

IIDRReplEngine

Source

DatastoreSource

Mirror DB (Operational Data Store)

OLTPDB

IIDR ReplEngineMirror

DatastoreMirror

IIDR ReplEngineDWH

DatastoreDWH

BackendDWH DB

Staging Layer

Raw + Business Data Vault

Mart Layer


END USER SAMPLE QUESTIONS


Which vehicles or aggregates are documented incompletely? (Data quality)

Which vehicles / which control units require SW updates?

Which interiors are most common by region?

Supply data for external simulations, customs clearance, spare part planning, etc.


Review the presented data warehouse architectures.

Which architecture would you recommend for

• An online store with real/right-time data integration needs

• Marketing department of a bank

List advantages and drawbacks of your proposal.

EXERCISE: RECOMMEND AN ARCHITECTURE


An online store with real-time/right-time data integration needs

• Architecture: Data Vault 2.0

• + Integration of many internal and external source systems (e.g. integrate social media data about the online store)

• + Fast data delivery in Raw Vault Layer (Real-time/Right-time data integration). Complex data cleansing / transformation / soft rules are delayed downstream towards Mart Layer

• - Transformation overhead (Source system data model to Data Vault data model to Dimensional data model)



Marketing department of a bank

• Architecture: Kimball Bus architecture

• + Start small for a department. If other departments are interested, new data and new Marts can be added on demand

• - High risk to loose the Enterprise view and several DWHs are built

That’s still quite a common scenario nowadays. A single Enterprise DWH is often not achieved (e.g. Mergers & Acquisitions, inflexibility due to a single centralized DWH, rapidly changing conditions, etc.) and therefore very often several DWHs with different architectures exist in parallel within a company.



• Now imagine that you prepare an exam.

• Identify 1-3 questions about DWH architecture (and/or DWH introduction) that you would ask in an exam.

• Write down the questions on stick-it cards.

EXERCISE - INTRODUCTION AND DWH ARCHITECTUREGROUP TASK


Which layers does the logical standard architecture have?

• Staging (Input), Integration (Cleansing), Core Warehouse (Storage), Aggregation, Mart (Reporting, Output) and additionally Metadata, Security, DWH Manager, Monitor

Which other architectures exist?

• Kimball Bus Architecture (Central data warehouse based on data marts)

• Inmon Corporate Information Factory

• Data Vault 2.0 Architecture (Dan Linstedt)

• DW 2.0: The Architecture for the Next Generation of Data Warehousing

• Operational Data Store (ODS)

SUMMARY DWH ARCHITECTURES


BIG DATA ARCHITECTURES


LOGICAL STANDARD DATA WAREHOUSE ARCHITECTURE


Data Warehouse

FrontendBackend




OLTP

OLTP

Core Warehouse

Layer(Storage

Layer)


(Reporting Layer)

Integration Layer

(Cleansing Layer)

Aggregation Layer

Metadata Management

Security


WHAT IS A DATA LAKE?


Hadoop

Dump anything in and wait?

Hoard 100ths ofPetabyte in HDFS?

Data Warehouse and Big Data / DHBWDaimler TSS 46

Data Lake on Hadoop

Data Swamp

Data Reservoir

Landing Zone

Data Library

Data Repository

Data Archive

Data Lake on Spark

Data Lake 3.0

IT‘S VERY HARD TO GET SPEED AND QUALITY [MARK MADSEN]


Schema-on-write• RDBMS: create data model firstSchema-on-read• Hadoop HDFS / NoSQL: create

data model later (when reading data)

RDBMS can also work with schema-on-read. Hadoop can also work with schema-on-write.

Dump question, but actually

there are many comparisons like that at the moment

• Hadoop is a tool / technology (or even many tools) like a RDBMS

• DWH is an architecture and concept• Architecture is abstraction and

defines a goal

• Architecture vs tools / technology

WHAT ARE DIFFERENCES BETWEEN HADOOP AND DWH?


• The data is not always known in advance, so it can’t be modeled in advance. [data can be anywhere → collect everything approach]

• The data architecture must be read-write from both the back and front, not a one-way data flow. The data written back may be repeatedly used, persistent data, or it may be temporary.

• The data may arrive with any frequency, and the rate may not be under your control.

ASSUMPTIONS + REQUIREMENTS HAVE CHANGED


• Naive idea: dump everything in (“landing zone”)

• Data hoarding is not a data management strategy

• A Data Lake brings in structure

• e.g. create directories in HDFS if Hadoop is used

• /transient

• /raw

• /standardized

• /use-case specific

HOW DOES A DATA LAKE DIFFER FROM A DATA SWAMP?


WHAT IS A DATA LAKE ACCORDING TO MARK MADSEN?SEPARATE COLLECT / MANAGE / DELIVERY


ZONES INSTEAD OF LAYERS ACCORDING TO MARK MADSEN


New data of unknown value, simple requests for new data can land here first, with little work by IT. Typically schema-

on-read.

More effort applied to data management: cleansed,

curated. Typically schema-on-write.

Optimized for specific uses / workloads. Personal folders.

Schema-on-write.

SCHEMA-ON-WRITE VS SCHEMA-ON-READ REVISITED


Old approach New approach

Model Collect

Collect Model

Analyze Analyze

Promotemove data into next

zone if required (datahas been analyzed andis needed more often)

DATA VAULT 2.0 ARCHITECTURE – TODAY’S WORLD (DANLINSTEDT)

Data Warehouse / DHBWDaimler TSS

https://www.youtube.com/watch?v=tDNjI1Yvqxw

DEFINING A DATA LAKE … BY DAN LINSTEDT



https://www.youtube.com/watch?v=tDNjI1Yvqxw

https://www.youtube.com/watch?time_continue=48&v=b9YfsjEjVS8

INTERVIEW WITH “FATHER OF DATA WAREHOUSING” BILL INMON



https://www.forestrimtech.com/video-talks-to-array-publications/

DATA LAKE TURNED INTO DATA SWAMP


Source: Ungerer: Cleaning Up the Data Lake with an Operational Data Hub, O’Reilly Media 2018, p.12

DATA LAKE


Source: Ungerer: Cleaning Up the Data Lake with an Operational Data Hub, O’Reilly Media 2018, p.11

• No agreed definition

• Characteristics [Madsen]:• Deals with data and schema

change easily

• Does not always require up front modeling

• Does not limit the format or structure of data

• Assumes a full range of data latencies, from streaming to one-time bulk loads, both in and out including write-back

Two different styles of work: different zones with different degrees ofgovernance

• Stability & Quality• Predictable, improving and

renovating

• Flexibility & Speed

• Explanatory

BIMODAL DATA GOVERNANCE


Source: https://www.gartner.com/it-glossary/bimodal/


https://www.gartner.com/it-glossary/bimodal/

• No agreed, standardized definition

• Additionally, there are many more buzzwords like Landing Zone, Data Repository, Data Swamp,

• Characteristics of a Data Lake architecture according to Madsen:

• Deals with data and schema change easily

• Does not always require up front modeling

• Does not limit the format or structure of data

• Assumes a full range of data latencies, from streaming to one-time bulk loads, both in and out including write-back

• Supports different uses of the same data



• A physical implementation of data solutions in the Hadoop ecosystem.

• A reservoir of curated, and often unconnected, datasets for data science, data exploration and reference data management use.

• The new Operational Data Store, a pub flap to hide a proper data management architecture.

• A name to indicate the staging area or the persistent staging area of a data warehouse solution.

• A technical solution to process huge volumes of data, for which an RDBMS is often too expensive to deploy.

THE (AB)USE OF THE WORDS „DATA LAKE“TECHNICAL


Source: https://www.linkedin.com/pulse/abuse-words-data-lake-martijn-ten-napel/


https://www.linkedin.com/pulse/abuse-words-data-lake-martijn-ten-napel/

• A concept that will spray fairy dust upon data management issues that arise from organisational issues, politics or a lack of understanding what working with data requires of an organisation.

• A term to use when you are really out of your comfort zone as an architect and to shelter behind.

• A clever marketing and sales trick, used to sell you a “must have” solution next to what you already have. You are told that without a data lake you will be unable to do data science, digital transformation or business analytics.

THE (AB)USE OF THE WORDS „DATA LAKE“CYNICAL REALITY


Source: https://www.linkedin.com/pulse/abuse-words-data-lake-martijn-ten-napel/


https://www.linkedin.com/pulse/abuse-words-data-lake-martijn-ten-napel/

SPECIFIC BIG DATA ARCHITECTURES

• There exist well-known reference architectures for Data Warehouses

• Many tools and schema-on-read came with the Hadoop ecosystem• Was a “black box” at the beginning

• Gets more and more structure with different layers instead of a “black box”

• Structure, modeling, organization, governance instead of tool-only focus


• Architecture by Nathan Marz

• Realtime and batch processing

• Batch layer stores and historizes raw data

• Serving layer contains batch views

• Query unions serving and realtime layer

• Rather complex

• Author recommends graph data model and advises against schema-on-read

LAMBDA ARCHITECTURE


• Lack of human fault tolerance

• Bugs will be deployed

• Operational errors will happen, e.g. accidentally delete data

• Data loss / corruption is worst case scenario

• Without data loss / corruption, mistakes can be fixed if original data is still available

• Must design for human error like you’d design for any other fault

WHAT IS THE MAIN INEVITABLE PROBLEM IN DATA SYSTEMS?


WHAT IS THE LAMBDA ARCHITECTURE?


Batch Layer

All Data (Master data set)

Speed Layer

RealTime Views

Serving Layer

Batch Views

Query(merge)

Data Stream

• Atomic, immutable data

• Ensure that data can not be deleted (accidentally)

• Fundamentally simpler

• Easy to implement on top of a distributed filesystem, eg Hadoop

• CR instead of CRUD• No updates

• No deletes

• Create (insert) and read (select) only

Immutability restricts the range of errors that can cause data loss/corruption

IMMUTABLE DATAMASTER DATA SET


NATHAN MARZ VIEW ON SCHEMA


• RawnessStore the data as it is. No transformations.

• ImmutabilityDon’t update or delete data, just

add more.

• Graph-like schema recommended

„Many developers go down the path of writing their raw data in a schemaless

format like JSON. This is appealing because of how easy it is to get started, but this

approach quickly leads to problems. Whether due to bugs or misunderstandings

between different developers, data corruption inevitably occurs“

(see page 103, Nathan Marz, „Big Data: Principles and best practices of scalable

realtime data systems", Manning Publications)

Source image: Nathan Marz, James Warren: Big Data: Principles and best practices of scalable realtime data systems, Manning Publications 2015

“My own personal opinion is that data analysis is much less important than data re-analysis. It’s hard for a data team to get things right on the very first try, and the team shouldn’t be faulted for their honest efforts. When everything is available for review, and when more data is added over time, you’ll increase your chances of converging to someplace near the truth.”–Jules J. Berman.

ANALYTICS VS RE-ANALYTICS


Source: http://www.odbms.org/blog/2014/07/big-data-science-interview-jules-j-berman/

http://www.odbms.org/blog/2014/07/big-data-science-interview-jules-j-berman/

NEW APPROACH TO BUILD DATA SYSTEMS


Immutable data(Source of Truth / Single version of

facts)

(Materialized) Views on data





Query = Application

HDFS / NoSQLRDBMSNewSQL






Query = Application


BATCH VIEWS AND REALTIME VIEWS


DATA ABSOBTION


https://dzone.com/articles/lambda-architecture-with-apache-spark

LAMBDA ARCHITECTURE – DATA EXAMPLECOMPUTE FOLLOWER LIST


Batch Layer

1.1. insert Jim1.1. insert Anne2.1 remove Jim3.1. insert George5.1. insert John (now)

Speed Layer

insert John (now)

ServingLayer

AnneGeorge

Query

Data Stream

AnneGeorgeJohn

Follower: 3

Computing best answer in real time may not always be possible

• Can compute exact answer in batch layer and approximate answer in realtime layer

• Best of both worlds of performance and accuracyFor example, a machine learning application where generation of the batch model requires so much time and resources that the best result achievable in real-time is computing and approximated updates of that model. In such cases, the batch and real-time layers cannot be merged, and the Lambda architecture must be used.

EVENTUAL ACCURACY


WHICH TOOLS COULD BE USED IN THE LAMBDAARCHITECTURE?


Batch Layer

All Data

Speed Layer

RealTime Views

Serving Layer

Batch Views

Query(merge)

Data Stream


CLOUD VENDORS OFFERINGS, E.G.AMAZON AWS


https://aws.amazon.com/de/blogs/big-data/unite-real-time-and-batch-analytics-using-the-big-data-lambda-architecture-without-servers/


https://aws.amazon.com/de/blogs/big-data/unite-real-time-and-batch-analytics-using-the-big-data-lambda-architecture-without-servers/

CLOUD VENDORS OFFERINGS, E.G.MICROSOFT AZURE


https://social.technet.microsoft.com/wiki/contents/articles/33626.lambda-architecture-implementation-using-microsoft-azure.aspx


https://social.technet.microsoft.com/wiki/contents/articles/33626.lambda-architecture-implementation-using-microsoft-azure.aspx


sou

rce:

Mar

kus

Sch

mid

ber

ger

-B

ig D

ata

ist

tot

–Es

leb

eB

usi

nes

s In

telli

gen

z?,

TDW

I 2

01

6Not relevant for exam


sou

rce:

Mar

kus

Sch

mid

ber

ger

-B

ig D

ata

ist

tot

–Es

leb

eB

usi

nes

s In

telli

gen

z?,

TDW

I 2

01

6

Lambda @glomex• Enrich batch-

driven data processing with real-time requirements

• Adapt Lambda architecture to own requirements

Remark: AWS lambda # lambda architecture


LAMBDA ARCHITECTURE – PROS AND CONS


Pro Con

Architecture emphasizes to keep data immutable. Mistakes can be corrected via recomputation

Maintaining code that needs to produce the same result in two complex distributed systems

Reprocessing is one of the key challenges of stream processing but is very often ignored

Operational burden of running and debugging two systems

• Incoming data is sent to batch and speed layer

• Batch layer constantly (re-) computes batch views

• Master data is stored in the batch layer in raw format: immutable & append-only

• Contains data except most up-to-date data due to high latency

• Replaces data in speed layer as soon as data in newer compared to speed layer

• Speed layer uses incremental algorithms to refresh real-time Views• Receives data stream for real-time processing

• Contains most up-to-date data only

• Serving layer contains views on batch layer data

• Merge is done in the query or in the serving layer

SUMMARY LAMBDA ARCHITECTURE


Jay Kreps wrote an article about „Questioning the Lambda architecture“

• He wrote about his experience with the Lambda architecture

• It works, but not very pleasant or productive

• Keeping code in sync is really hard

• Need to build complex, low-latency processing systems

• Scalable high-latency batch system

• Low-latency stream stream-processing system

• Instead of duct taping batch & speed:

→ Kappa architecture / Log-centric architecture / Stream data platform

QUESTIONING THE LAMBDA ARCHITECTURE


Source: https://www.oreilly.com/ideas/questioning-the-lambda-architecture

https://www.oreilly.com/ideas/questioning-the-lambda-architecture

• Architecture by Jay Kreps

• Logcentric, write-ahead logging

• Each event is an immutable log entry and is added to the end of the log

• Read and write operations are separated

• Materialized views can be recomputed consistently from data in the log

KAPPA ARCHITECTURE


Lambda architecture

• Complex

• It works, but not very pleasant or productive: keeping code in sync is really hard

• Addresses the challenge of reprocessing

• Batch views to contain updated data (a Lambda architecture rcommendation)

• Bugs (true requirement for reprocessing)

• New user requirements (true „requirement“ for reprocessing)

Logs unify• Batch processing + Stream processing

• In Kappa, reprocessing required only when processing logic has been modified

REPROCESSING


KAPPA ARCHITECTURE


Source: https://www.oreilly.com/ideas/questioning-the-lambda-architecture

https://www.oreilly.com/ideas/questioning-the-lambda-architecture

SUMMARY: ENTERPRISE-WIDE ARCHITECTURE


Source: Jay Kreps: I heart logs, O’Reilly 2014

• The LDW is a multi-server / multi-engine architecture

• The LDW can have multiple engines, the main ones being: The data warehouse(DW), data marts and the data lake

• We should not be trying to choose a single engine for all our requirements. Instead, we should be sensibly distributing all our requirements across the various components we have to choose from. Reasons:

• Regulatory constraints

• Organizational constraints

LOGICAL ARCHITECTURE (GARTNER)


https://blogs.gartner.com/henry-cook/2018/01/18/logical-data-warehouse-project-plans/


https://blogs.gartner.com/henry-cook/2018/01/18/logical-data-warehouse-project-plans/

VIRTUAL ARCHITECTURE


Data Warehouse

FrontendBackend



Data Lake

OLTPQuery Management

Weakly+partly subject-oriented, Weakly+partly integrated,

Not time-variant,Not non-volatile

Views,Alias names,

Metadata

DWH


• Data not extracted from operational systems and stored separately

• Standardized interface for all operational data sources

• One "GUI" for all existing data

• Generates combined queries

• Query Processor joins query result data from different sources

• Can also access data in Hadoop (Polybase, Big SQL, BigData SQL, etc)




• Query Management manages metadata about all operational systems

• (physical) location of data and algorithms for extracting data from OLTP system

• Implementation easier

• Low cost: can use existing hardware infrastructure

• Queries cause significant performance problems in operational systems

• Known problems when analyzing operational data directly

• Same query is processed multiple times (if queried multiple times)

• Same query delivers different results when processed at different times




DATA MANAGEMENT ACCORDING TO BARCFLEXIBILITY CAUSES HETEROGENITY


Source: https://www.datafestival.de/events/proof-of-concepts-als-strategietool-erfahrungen-aus-der-medienbranche-de/; Jacqueline Bloemen, Datenarchitektur für Business Analytics – was Sie berücksichtigen sollten, DataFestival 2019


https://www.datafestival.de/events/proof-of-concepts-als-strategietool-erfahrungen-aus-der-medienbranche-de/

Daimler TSS GmbHWilhelm-Runge-Straße 11, 89081 Ulm / Telefon +49 731 505-06 / Fax +49 731 505-65 99

[email protected] / Internet: www.daimler-tss.com/ Intranet-Portal-Code: @TSSDomicile and Court of Registry: Ulm / HRB-Nr.: 3844 / Management: Christoph Röger (CEO), Steffen Bäuerle


THANK YOU

USE CASE: ANALYSIS BATTERY AGING


Max capacityCurrent capacity

• JSON data ingested into HDFS, Hive tables on JSON files

• Identify breaks (“> 8h”) and compute current drain

• Sensor data format change without notice

• Sensors get regularly updated with new versions

• Names of metrics may change

• Sensors with various versions in the field

• Sensors from different suppliers

• Often many fields >>100 and increasing with new sensor versions

• Easy storing of data in HDFS and applying schema later

• Data from Robots, vehicles, …

STRUCTURING THE DATA LAKENEW DATA SOURCES – SENSOR DATA


• Sensor data format change without notice• Time consuming and error-prone

data integration into the Data Lake

• Therefore preparation of data for usage in the Data Reservoir required: “Data Engineer”

STRUCTURING THE DATA LAKE“SCHEMA-ON-READ”


Raw dataD

ata

Go

vern

ance

Consumption

Enhanced data

Met

adat

a M

anag

eme

nt

Data A

rchival

Data Secu

rity

json

Samp-ling / filter

Hive tables

Hive tables

Struc-ture

R Python

A holding of 3 telecommunication companies

• Architecture: Virtual Data Warehouse

• + Companies may not want to provide their data to a new storage

• + Can easily be extended if new companies join the holding or reduced if a company leaves the holding

• - Bad performance

• - Not really data integration achieved, low Data Quality

• - Firewalls have to be opened



WHERE TO USE HADOOP AND SPARK?


Source: Rick F. van der Lans: New Data Storage Technologies, TDWI Munich 2018

WHERE TO USE NOSQL?


Source: Rick F. van der Lans: New Data Storage Technologies, TDWI Munich 2018

Apple: multiple Petabytes

• Customer insights: who’s who and what are the customers up to

Walmart: 300TB (2003), several PB today

• It tells suppliers, “You have three feet of shelf space. Optimize it.”

eBay: >10PB, 100s of production DBs fed in

• Get better understanding of customers

Most DWHs are much smaller though. For huge and small DWHs: High challenges to architect + develop + maintain + run such complex systemshttps://gigaom.com/2013/03/27/why-apple-ebay-and-walmart-have-some-of-the-biggest-data-warehouses-youve-ever-seen/ and http://www.dbms2.com/2009/04/30/ebays-two-enormous-data-warehouses/

EXAMPLES OF DATA WAREHOUSES IN THE INDUSTRY


https://gigaom.com/2013/03/27/why-apple-ebay-and-walmart-have-some-of-the-biggest-data-warehouses-youve-ever-seen/

http://www.dbms2.com/2009/04/30/ebays-two-enormous-data-warehouses/

• Architecture, conceptData Lake

• Tools (that can be used to implement a Lake)

Hadoop, Spark, Elastic Stack

DATA LAKE VS HADOOP


DWH AND DATA LAKE


DWH on RDBMS

Slowly Changing DimensionELT vs ETL3-Layer vs 2-LayerKimball ApproachInmon DefinitionStar SchemaData VaultAnchor Modelingetc

Data Lake on Hadoop

Schema-on-ReadAgilityParquetHiveHBaseSQL-on-HadoopImpalaOozieZoekeeper

Methods, Concepts,

Techniques

Tools,Tools,Tools

DATA LAKE (MARTIN FOWLER)


Source: https://martinfowler.com/bliki/DataLake.html

https://martinfowler.com/bliki/DataLake.html

A Data Lake acquires data from multiple sources in an enterprise in its native form and may also have internal, modeled forms of this same data for various purposes. The information thus handled could be any type of information, ranging from structured or semi-structured data to completely unstructured data. A Data Lake is expected to be able to derive enterprise-relevant meanings and insights from this information using various analysis and machine learning algorithms.



Source: Pankaj Misra, Tomcy John: Data Lake for Enterprises Packt 2017

DATA LAKE LIFE CYCLE


Source: Pankaj Misra, Tomcy John: Data Lake for Enterprises Packt 2017

DATA LAKE (ECKERSON GROUP)


Source: https://www.eckerson.com/articles/ten-characteristics-of-a-modern-data-architecture

https://www.eckerson.com/articles/ten-characteristics-of-a-modern-data-architecture

LAMBDA ARCHITECTURE


Source:

Batch Layer

BatchEngine Serving Layer

ServingBackend Queries

Raw historydata

Resultdata

Real-Time Layer

Real-TimeEngine

OVERVIEW LAMBDA ARCHITECTURE


https://de.slideshare.net/gschmutz/big-data-and-fast-data-lambda-architecture-in-action

https://de.slideshare.net/gschmutz/big-data-and-fast-data-lambda-architecture-in-action

WHICH TOOLS COULD BE USED IN THE LAMBDAARCHITECTURE? DB REQUIREMENTS


Easier to implement fordatabase vendors

compared to randomaccess

Batch layer

Write sequential

once

Bulk sequential read many

times

Speed layer

Random write

Random read

Servinglayer

Batch write

Random read

WHICH TOOLS COULD BE USED IN THE LAMBDAARCHITECTURE? DB REQUIREMENTS


More challenging

More challenging More challenging

WHAT ARE THE THREE PARADIGMS OF PROGRAMMING?

• Request/Response

• Batch

• Stream processing


Source: https://de.slideshare.net/JayKreps1/distributed-stream-processing-with-apache-kafka-71737619

https://de.slideshare.net/JayKreps1/distributed-stream-processing-with-apache-kafka-71737619

WHAT IS A LOG?



• Database transactions/data

• User, products, etc.

• Events• Tweets, clicks, impressions, pageviews, etc.

• Application metrics

• CPU usage, requests, etc.

• Application logs• Service calls, errors, etc.

TYPES OF DATA


• A bank account’s current balance can be built from a complete list of its debits and credits, but the inverse is not true.

• In this way, the log of transactions is the more “fundamental” data structure than the database records storing the results of those transactions.

A software application’s database is better thought of as a series of time-ordered immutable facts collected since that system was born, instead of as a current snapshot of all data records as of right now.

WHAT IS A LOG?


Source: https://blog.parse.ly/post/1550/kreps-logs/

https://blog.parse.ly/post/1550/kreps-logs/

ENTERPRISE-WIDE ARCHITECTUREUSING LOG-CENTRIC APPROACH


Source: https://www.confluent.io/blog/using-logs-to-build-a-solid-data-infrastructure-or-why-dual-writes-are-a-bad-idea/

https://www.confluent.io/blog/using-logs-to-build-a-solid-data-infrastructure-or-why-dual-writes-are-a-bad-idea/

ENTERPRISE-WIDE ARCHITECTUREUSING LOG-CENTRIC APPROACH AND CDC




ENTERPRISE-WIDE ARCHITECTURE TODAYDATA INTEGRATION CHALLENGE




LOG CENTRIC ARCHITECTURE / KAPPA ARCHITECTURE / DATA STREAMING PLATFORM



Most common understanding (doing the „T“ in ETL):

• Stream processing is the parallel processing of data in motion = computing on data directly as it is produced or received.

• Not necessarily transient, approximate, lossy (assumptions from Lambda architecture and other event processing systems)

WHAT IS STREAM PROCESSING?


• Active MQ, RabbitMQ

• Problems:• Not distributed

• Throughput

• Persistence

• Ordering

MESSAGING SYSTEMS @LINKEDIN

1st attempt

• Kafka

• Key abstraction: Logs

• build from scratch• Distributed system by design

• Partitioning with local ordering

• Elastic scaling

• Fault tolerance

2nd attempt


• Scalability

• Hundreds of MB/sec/server throughput

• Many TB per node

• Guarantees of a database

• All messages strictly ordered (within a partition)

• All data persistent

• Distributed by default

• Replication

• Partitioning

• Producers + consumers all fault tolerant and horizontally scalable

KAFKA: A MODERN DISTRIBUTED SYSTEM FOR STREAMS


ETL REVISITED WITH KAFKA CONNECT AND KAFKA STREAMS


Source: https://de.slideshare.net/JayKreps1/distributed-stream-processing-with-apache-kafka-71737619

https://de.slideshare.net/JayKreps1/distributed-stream-processing-with-apache-kafka-71737619

• Apache Spark streaming

• Apache Storm (twitter, Nathan Marz)

• Apache Samza (linkedin)

• Apache Flink

• Apache Kafka Streams

• Simple library

• Reprocessing

• No microbatch = everything is a stream

• Local state

• Key operations: filter, aggregate, join

WHAT ARE STREAM PROCESSING FRAMEWORKS?


KAFKA @LINKEDIN


Source: https://de.slideshare.net/JayKreps1/i-32858698

https://de.slideshare.net/JayKreps1/i-32858698

SUMMARY: STREAM DATA PLATFORM USING KAFKAUNIFYING BATCH AND STREAM PROCESSING


Source: https://de.slideshare.net/JayKreps1/i-32858698

https://de.slideshare.net/JayKreps1/i-32858698

• A message-oriented implementation requires an efficient messaging backbone that facilitates the exchange of data in a reliable and secure way with the lowest latency possible.

• Creating small, self-contained, data-driven applications that meld streaming data and microservices together is a good practice to break down large problems and projects into approachable chunks, reduce risk, and deliver value faster.

• Think of combinations of data-processing applications with microservices to deliver specific features and insights from a data stream.

KAFKA AND MICROSERVICES


KAPPA ARCHITECTURE


Source:

Real-Time Layer

Real-TimeEngine

Serving Layer

ServingBackendData Queries

Raw historydata

Resultdata

PS-3C ARCHITECTURE


Source: TDWI 2016

Data Library

StorageEngine

3C Layer

PreparationEngineData Queries

Serving Layer

DeliveryEngine

Raw historydata

Integrated subjects

Resultdata

• Architecture by Rogier Werschkull

• Store incoming data in Data Library Layer (Persistent staging = PS)

• Prepare data in a 3C layer for “Concept – Context – Connector”-model

• Concept + Connector can be virtualized on data in Data Library Layer

PS-3C ARCHITECTURE


• Architecture by Joe Caserta

• Big Data Warehouse may live in one or more platforms on premises or in the cloud

• Hadoop only

• Hadoop + MPP or RDBMS

• Additionally NoSQL or Search

POLYGLOT WAREHOUSE


POLYGLOT WAREHOUSE


Source: https://www.slideshare.net/CasertaConcepts/hadoop-and-your-data-warehouse

• Architecture by Claudia Imhoff

• combine the stability and reliability of the BI architectures while embracing new and innovative technologies and techniques

• 3 components that extend the EDW environment

• Investigative computing platform

• Data refinery

• Real-time (RT) analysis platform

THE EXTENDED DATA WAREHOUSE ARCHITECTURE (XDW)THE ENTERPRISE ANALYTICS ARCHITECTURE


THE EXTENDED DATA WAREHOUSE ARCHITECTURE (XDW)THE ENTERPRISE ANALYTICS ARCHITECTURE


Source: https://upside.tdwi.org/articles/2016/03/15/extending-traditional-data-warehouse.aspx

• Inflow Lake: accommodates a collection of data ingested from many different sources that are disconnected outside the lake but can be used together by being colocated within a single place

• Outflow Lake: a landing area for freshly arrived data available for immediate access or via streaming. It employs schema-on-read for the downstream data interpretation and refinement.

• Data Science Lab: most suitable for data discovery and for developing new advanced analytics models

GARTNER DATA LAKE ARCHITECTURE STYLES


Source: http://blogs.gartner.com/nick-heudecker/data-lake-webinar-recap/

GARTNER DATA LAKE ARCHITECTURE STYLES


Source: http://blogs.gartner.com/nick-heudecker/data-lake-webinar-recap/

GARTNER – THE LOGICAL DWH


https://blogs.gartner.com/henry-cook/2018/01/28/the-logical-data-warehouse-and-its-jobs-to-be-done/

https://blogs.gartner.com/henry-cook/2018/01/28/the-logical-data-warehouse-and-its-jobs-to-be-done/

SUMMARY


Landing Area

StorageEngine

Data Lake

IntegrationEngineData Queries

Data Presentation

DeliveryEngine

Raw historydata

Lightly integrated data

Resultdata

• Architecture by Eckerson Group

• DWHs exist together with MDM, ODS, and portions of the data lake as a collection of data that is curated, profiled, and trusted for enterprise reporting and analysis

DATA CORE – DAVE WELLS


https://www.eckerson.com/articles/the-future-of-the-data-warehouse

DATA CORE – DAVE WELLS



DATA LIFECYCLE – DAVE WELLS



DWH AND DATA LAKE – DAVE WELLSIN PARALLEL VS INSIDE



• Dimensional modeling is not dead.

• The benefits are still valid in the age of Big Data, Hadoop, Spark, etc:

• Eliminate joins

• Data model is understandable for end users

• Well-suited for columnar storage + processing (e.g. SIMD)

• Nesting technique

• E.g. tables with lower granularity can be nested into large fact table

• Usage in SQL: Flatten(kvgen(<json>))

DIMENSIONAL MODELING IN THE AGE OF BIG DATA


SELF-SERVICE DATA


https://www.oreilly.com/ideas/how-self-service-data-avoids-the-dangers-of-shadow-analytics

Functional area Why important Self-service approach

Data acceleration

With shadow analytics, users create redundant data copies.

The system must be capable of autonomously identifying the best optimizations and adapting to emerging query patterns over time.

Data catalog Data consumers struggle to find data that is important to their work. Users keep private notes about data sources and data quality, meaning there is no governance.

In the self-service approach, the catalog is automatic—as new data sources are brought online.

Data virtualization

It is virtually impossible for an organization to centralize all data in a single system.

Data consumers need to be able to access all data sets equally well, regardless of the underlying technology or location of the system.

SELF-SERVICE DATA




Data curation There is no single “shape” of data that works for everyone.

Data consumers need the ability to interact with data sets from the context of the data itself, not exclusively from simple metadata that fails to tell the whole story. Data consumers should be capable of reshaping data to their own needs without writing any code or learning new languages.

Data lineage As data is accessed by data consumers and in different processes, it is important to track the provenance of the data, who accessed the data, how the data was accessed, what tools were used, and what results were obtained.

As users reshape and share data sets with one another through a virtual context, a self-service data platform can seamlessly track these actions and all states of data along the way, providing full audit capabilities as well.

SELF-SERVICE DATA




Open source Because data is essential to every area of every business, the underlying data formats and technologies used to access and process the data should be open source

Self-service data platforms build on open source standards like Apache Parquet, Apache Arrow, and Apache Calcite to store, query, and analyze data from any source.

Security controls

Organizations safeguard their data assets with security controls that govern authentication (you are who you say you are), authorization (you can perform specific actions), auditing (a record of the actions you take), and encryption (you can only read the data if you have the right key).

Self-service data platforms integrate with existing security controls of the organization, such as LDAP and Kerberos.

Hadoop (and for similar reasons Spark) has its strengths but no as a DWH replacement, e.g.

• Fast query reads only possible in HBase with an inflexible (use case specific) data model

• No sophisticated query optimizer

• Hadoop is very complex with many tools/versions/vendors and no standard

• Security is still at the beginning

IS THE DWH DEAD?


Digitization is the process of converting information into a digital (i.e. computer-readable) format

DIGITIZATION – THE DIGITAL DATA EXPLOSION


Source: https://slideplayer.com/slide/10254426/

https://slideplayer.com/slide/10254426/

FUTURE MOBILITY


https://blog.daimler.com/en/2018/07/19/future-mobility-metropolitan-cities-congress/

https://blog.daimler.com/en/2018/07/19/future-mobility-metropolitan-cities-congress/

Display video

IS CHINA THE NEXT SILICON VALLEY?


SOME STATISTICSBAT – BAIDU + ALIBABA + TENCENT


Sources: https://venitism.wordpress.com/2017/12/15/beware-of-the-bats-baidu-alibaba-and-tencent/https://www.afr.com/brand/business-summit/baidu-alibaba-tencent-to-disrupt-facebook-amazon-netflix-google-in-asia-20180228-h0wrdl

https://venitism.wordpress.com/2017/12/15/beware-of-the-bats-baidu-alibaba-and-tencent/

https://www.afr.com/brand/business-summit/baidu-alibaba-tencent-to-disrupt-facebook-amazon-netflix-google-in-asia-20180228-h0wrdl

VolumeVelocityVarietyVeracity

Value

DIGITIZATION – CHALLENGES & OPPORTUNITIES


Ethics

AI

Impact

IoT, Industry 4.0, …

Data integration

Real-time

Data quality

Data-driven

Digital Services

DIGITIZATION PUBLIC AUTHORITIES


lecture @dhbw: data warehouse 02 dwh and big data...

Documents