BI on Big Data Presentation

BI on Big Data
Strata San Jose, March, 2018
Shant Hovsepian, Arcadia Data
Mark Madsen, Think Big Analytics

Users hate BI. And have for a long time

The old problem was access, the new problem is analysis

You keep using that word. I
do not think it means what
you think it means.
What do you mean by “analytics”?

User-focused criteria: context and point of use
Information use is diverse and
varies based on context:
▪ Get a quick answer
▪ Solve a one-off problem
▪ Analyze causes
▪ Do experiments
▪ Make repetitive decisions
▪ Use data in routine processes
▪ Make complex decisions
▪ Choose a course of action
▪ Convince others to take
action
One size doesn’t fit all.

There are two parts to “analytics”
The mathy stuff The query & reporting stuff

Analysis: the verb that’s ignored
Analysis is something people do across this spectrum
Good paper on the topic of analysis: Enterprise Data Analysis and Visualization: An Interview Study
http://vis.stanford.edu/files/2012-EnterpriseAnalysisInterviews-VAST.pdf

Old market says: There’s nothing wrong with what
you have, just keep buying new products from us

The big data market has an answer…

The data lake: just dump the data in!

Combine
with self-
service tools:
we’ll figure it
out later!

The primary view of BI, self service is publishing data

An architectural history of BI tools
First there were files and reporting programs.
Application files feed through a data processing pipeline
generate an output file. The file is used by a report
formatter for print/screen.
Every report is a program written by a developer.

We had the concept of cubes before we had RDBMSs.
First commercial product (Express) in 1970.
Provided interactive response time but had problems:
rigidly defined schema, inflexible definition, data size
limits, slow cube build times, resulting in cube explosions.

Then we had databases and tools with embedded SQL,
and query-by-example templating soon after.
These had better scalability and flexibility over OLAP
models. They decoupled storage from access from the
rendering of the interface.
SQL

With more regular schema models, in particular
dimensional models that didn’t contain cyclic join paths,
it was possible to automate SQL generation via semantic
mapping layers. Query via business terms made BI usable
by non-technical people.
SQL

Response time is a key driver for analysis
Many approaches today are a step backward. Unless you
resolve this task performance gap, real analysis work is a
challenge and will remain the domain of Excel and tools
that extract subsets of data into memory.
Days
Hours
Minutes
Seconds
Instantaneous
come back tomorrow
go to lunch
take a break
get some coffee
check email/FB
take a sip of coffee
immerse yourself in work
Flow is possible only in the “less
than 3 second” range

BI server architecture shifts
The SQL-generating server model of BI scales
extremely well but has poor user response time.
Solution 1: pre-cache
query results or prebuild
datasets on the BI server
(i.e. the old OLAP model)
Well-known problems
with this.
Solution 2: Shove all the
data into a BI server
repository. Avoids subset
problems. Adds potential
scaling problems.

IT reality today is multiple data stores, not one place
Independent, purpose-built databases and processing systems for
different types of data and query / computing workloads is the new
norm for information delivery. No single place for data or access.
BI, dashboards,
analytics, apps
1 MargeInovera $150,000 Statistician
2 AnitaBath $120,000 Sewerinspector
3 IvanAwfulitch $160,000 Dermatologist
4 NadiaGeddit $36,000 DBA
Query
processing
Databases Documents Flat Files Objects Streams ERP SaaS Applications
Source Environments
Data
processing
Stream
processing

There is always a third way
The previous choices were driven by client-server
thinking. We have a distributed (cloud) environment.
Possibilities:
Don’t force all the compute
into the DB or server.
Don’t force all the compute
to the client.
Data on demand, bring it to
the analysis from where it is,
and/or execute the analysis
local to where the data is.

Connectors to data sources the tool can
communicate with.
Semantic layer tools: what we’re used to
Client requests and responses in
native format
BI server:
Semantic layer
Query generation
Connectors
SQL and NoSQL
connectors
Note: most BI servers
still can’t talk to more
than one DB at the same
time

Map-based tools: what we’re back to
Internal or external
API-accessible sources.
Connector based data
sources the tool can
communicate with.
Query from client / server
Queries sources directly
Any integration done locally
No semantic layer translating
SQL and NoSQL
connectors
XML, JSON, text,
binary return
formats
Note: most products have minimal to
no local join ability

With analysis, the BI approach has to be inverted
The process used today for data warehousing:
1. Model
2. Collect
3. Analyze
The new process is:
1. Collect
2. Analyze
3. Model
This is a shift from
planned design to
adaptive design for data
management, and a
multi-skill team.

Analysis and BI: Discoverability and Repeatability
Discover
Explore
Analyze
Model
Consume
Promote
Focus on repeatability
Application cycle time
Focus on discoverability
Analyst cycle time
80% of data use 80% of analysis

Just enough modeling: an analogy for analytic data
Multiple contexts of use, differing quality levels
You need to keep the original because just like baking,
you can’t unmake dough once it’s mixed.

A core problem with one global schema is change

Big data answer? Schema on read
Prof. Dr. Jens Albrecht

Schema on read is an answer, sometimes
Schema-on-read is really only good for the developer who doesn’t
know what to do with the data. There is a price to pay with schema
on read, but you usually don’t see it at the beginning.
SoR problem: metadata is what you wished your data looked like.
Reality is not requirements = code
Reality is the data, not the metadata

How did we get to this
point with BI & big data?
There’s a difference
between having no past
and actively rejecting it.

Move Data to Separate BI Server
Move Data to BI Server
BI & Visualization Server
Pros
✓ Least Costly
✓ Use existing BI tools
Cons
✘ Shallow insights – summary data
✘ Requires IT/DBA: new views & data
movement
✘ Separate security models
✘ Not real-time: batch data updates
✘ Heaviest burden on network

ODBC to SQL in Big Data, SQL on Big Data
Pros
✓ Can get detailed data
✓ Performance leverages the architecture
Cons
✘ Lower user concurrency
✘ Cannot access unstructured data (requires
schema)
✘ Cost - Manage security in multiple tools,
separate administration for metadata
(Impala, Vertica, Hawq,
Presto)BI & Visualization Server
(SparkSQL, Hive, Athena)
BI & Visualization Server

(R)OLAP on Hadoop
Pros
✓ Use existing BI tools
✓ Higher user concurrency
Cons
✘ Lacks ad-hoc freedom - Requires IT/DBA for
new views
✘ Not real-time: batch data updates
✘ Cannot access unstructured data (requires
schema)
✘ Cost – Multiple tools and data duplication
✘ Increased administration – Separate security
models, administration
OLAP Middleware

Native BI
Pros
✓ Greatest user concurrency
✓ Linear scalability
✓ Agility for analysts (drill to detail)
✓ Supports complex data sources
✓ Real time
✓ Lowest TCO: simplified architecture
Cons
✘ Newer technology and approach
✘ Requires some Hadoop skills to set up and
maintainBI runs on Hadoop

Hadoop: it disaggregates the database
One of the key things Hadoop does is to separate the
storage, execution and API layers of a database. This
allows for processing flexibility, but it does not permit
one to build a reliable, high performance database
across the layers. You trade these for write flexibility.
Hadoop distributed filesystem (HDFS)
General-purpose data engines
Abstraction layers
Storage management

A more specific look at layers and engines
Base storage
SQL, MDX
Kylin
Storage mgmt
Engine
Abstraction
layer / API
You can program to any layer you
choose. Some projects build on top of
multiple others.
Language/API Engine
MapReduce Tez
Cascading
Spark
Storage (filetypes in HDFS, Hbase, etc)
Crunch
Pig
Hive
SparkSQL
NativeAPI
Giraph
Hive
Crunch
Pig
Impala
Drill
Presto
NativeAPI
NativeAPI
Hive
Pig
NativeAPI
Hbase
Phoenix

Four models for SQL on Hadoop
1. Parse and compile SQL into MapReduce jobs
2. Put a SQL interpreter on a generic execution engine
3. Run a native SQL engine in the cluster
4. Run a SQL interpreter on a non-generic engine (this will
limit SQL functionality based on the underlying engine).
MapReduce TezSpark
Storage (filetypes in HDFS, Hbase, etc)
Hive
SparkSQL
Hive
Impala
Drill
Presto
Hive
Druid
Hbase
Phoenix1 2 32 4
Metanautix
JethroDB

What’s under the hood matters in when querying
Source: Randy Bias

The core problem of old BI was scalability. This is solved.
New uses require new platforms for different workloads.
Source: Noumenal, Inc.
Note: this was 8 years
ago. Largest MPP
RDBMS I know is 99PB
Standard DBs
MPP DBs
Platforms

The shifting BI paradigm
The tool market is shifting,
driven by new architectures
that are enabled by new
technologies.
Front-end tools are evolving
away from BI-as-publishing,
which changes their design,
increases the burden on back
end databases and creates new
interaction challenges.
You need to evaluate tools
based on more usage scenarios
and interactive capabilities,
less on report-building /
dashboard features.
One standard tool is not the
norm, it’s the exception.

TANSTAAFL
When replacing the old
with the new (or ignoring
the new over the old) you
always make tradeoffs,
and usually you won’t see
them for a long time.
Technologies are not
perfect replacements for
one another. Often not
better, only different.

The right tool is the one that people will actually use,
not the one you want them to use
“But we already have an enterprise standard”

Mark Madsen is the global head of
architecture at Teradata, Prior to that
he was president of Third Nature, a
research and consulting firm focused
on analytics, data integration and data
management. Mark is an award-
winning author, architect and CTO
whose work has been featured in
numerous industry publications. Over
the past ten years Mark received
awards for his work from the American
Productivity & Quality Center, TDWI,
and the Smithsonian Institute. He is an
international speaker, chairs several
conferences, and is on the O’Reilly
Strata program committee. For more
information or to contact Mark, follow
@markmadsen on Twitter or visit
http://ThirdNature.net
About the Presenter

About the Presenter
Shant Hovsepian is a cofounder
and CTO of Arcadia Data, where he
is responsible for the company’s
long-term innovation and technical
direction. Previously, Shant was a
member of the engineering team
at Teradata, which he joined
through the acquisition of Aster
Data. Shant interned at Google,
where he worked on optimizing the
AdWords database, and was a
graduate student in computer
science at UCLA. He is the coauthor
of publications in the areas of
modular database design and high-
performance storage systems.

BI on Big Data Presentation

More Related Content

Similar to BI on Big Data Presentation (20)

More from Arcadia Data (13)

Recently uploaded (20)

BI on Big Data Presentation