1. Imagination at work
Industrial Internet Case Study
using HBase and TSDB
Shyam Varan Nath
Arnab Guin
May, 2015

2. Agenda
• Introduction to IoT and Industrial Internet
• Industrial Use Case
• Technology Details
• Wrap up

3. About Shyam
• Architect – Industrial Analytics @GE
• Worked in IBM, Deloitte, Oracle and Halliburton,
prior to GE
• Regular speaker in technology conferences such
as Oracle Openworld, IoT Summit, Collaborate,
BIWA Summit on IoT, Big Data and Analytics
related topics
• Education: IIT Kanpur (B Tech EE), MBA and MS
Computer Science from Florida Atlantic University

4. About Arnab
• Staff Software Engineer, Big Data – Predix
Platform @GE
• Working on big data, analytics (Past HBaseCon
attendee)
• Past work in several domains – audience
research, genomics, compilers, EE
• Education: BS Computer Science, Masters
(Quantitative Finance), CS/EE graduate studies
@Stanford

5. GE – Guiding Principle
Industrial Internet Makes Machines Better!

6. Overview – Industrial Internet as a Service
Industrial Data Lake
Industrial Internet Application
Industrial
Machines
https://www.gesoftware.com/news-events/featured-stories/controls%E2%80%94-brilliant-edge-ge%E2%80%99s-industrial-internet
Connectivity
Analytics

7. Value to Business User

8. Aviation Use Case - Jet Engines
Speed Sensor
Exhaust Gas Temperature
(EGT) sensors
Temperature Sensors
Temperature Sensors
Pressure sensors
* Simplified view of some of the sensors

9. Making “Sense” of the “Sensors”
EGT = Exhaust Gas Temperature
The temperature of the exhaust gases as they enter the tail
pipe, after passing through the turbine
A good indicator of the health of engine (just like human body
temperature)
Recording & interpreting the EGT can help to detect several
jet engine
problems.
http://www.geaviation.com/press/cf34/cf34_20140513b.html

10. Business Problem
• The aircraft collects data about it’s operations including the jet
engine using Quick Access Recorder (QAR)
• Engine analytics applied to such data in near real-time can be
used to proactively diagnose problems and reduce unplanned
downtime
• Continuous Engine Operations Data (CEOD) can be up to 500GB
per flight. With 300K flights per day, it soon becomes a Big Data
problem
• Data Lake is a fertile ground for Big Data Analytics to understand
jet engine behavior and problems over its age of ~30 yrs
• Analytics developed with full data, can be deployed to summary
information, near real-time
http://en.wikipedia.org/wiki/Quick_access_recorder

11. All data
Access to real-time data
and historical data and not
limited to snapshot of data
Any data
Handling of all data types
including documents, images
machine data, sensor data
One place
Access to all data in one
place to quickly respond to
the speed of business change
1
2
3
Rapid access to all data for analytics
How long will
it last without
failures or
maintenance?
Is my asset
ready when
there is market
opportunity?
Is my asset
performing
optimally?
How to
configure
for best
operational
results?
FLEXIBLE DATA MODELS
New approach – Industrial Data Lake
architecture
INDUSTRIAL DATA LAKE
Data scientist Field operations Business analyst
Sensor
data
Content
(images, videos,
manuals, etc.)
Machine
data
Historian
data
CRM,
ERP,
etc.
Logs,
click
streams
Geo-
location
data
Social
network
data

12. Industrial Data Lake
50BMachines will be
connected on the
internet by 2020
2XIndustrial data
growth within
next 10 years
*Source: IDC
CRM, ERP,
etc. Logs
Social
network
data
Geo-location
data
In practice only
3%of potentially useful
data is tagged
and even less
is analyzed*
9M
Data points
per hour for each
locomotive
500GB
Data per blade
by gas
turbines
Sensor
data
Content
(images, videos,
manuals, etc.)
Historian
data
Machine
data
35GB
Data per day
from each
Smart Meter
50X
Data growth
in healthcare
(2012 – 2020)
1TB
Data per
flight
© General Electric Company, 2014. All Rights Reserved.

13. Machine
data
Data Flow - Ingestion, Storage, Analytics
Sensor
data

14. System Components
Ingestion
- High speed real-time data input in the time domain
(streams)
- Batch processing (files)
Transport Layer
- RabbitMQ
Security
Fault tolerance
- Multiple containers (ingestion)
- Highly available queues (transport)
- Multiple masters, replication (storage)
- Multi-node zookeeper quorum (coordination)
Ingestion
Storage
Transport
Security
Zookeeper

15. Read-Write Tracks
W
R
I
T
E
H
B
A
S
E
T
S
D
B
TimeRange
Tag/Metric
Tag Value
Attributes
R
E
A
D
H
B
A
S
E
T
S
D
B
Timestamp
Tag/Metric
Tag Value
Attributes
TimeMessag
e
Block
Write
Atomic
Write
Async
Parallel
Async
ParallelTimeQuer
y
Block
Read
Atomic
Read
s
e0
now
time
Block reads
can be
syncronized
(futures)

16. ● Abstractions over TSDB for simplification
● Higher throughput reads
o Multi-threading
o Multi-processing (coprocessors) - wip
o Divide and conquer
o On-demand loading (yield to wield)
Read/Writ
e APIs KairosDB
OpenTSDB
HBase
Metric TimeStamp
……..
Salt
.
.
Region
Servers
APIs (Read + Write)
Hbase
DataStore
Plugin

17. Performance
Region
Servers
Higher throughput reads
● Multi-threading (client side)
● Multi-processing (coprocessors)
● Divide and conquer
● On-demand loading (read ahead and iterate)

18. Aggregation
Use cases
● Exploratory analysis (random/stratified sampling)
● Graphing/plotting data
● Trend analysis (regression)
dbObject.readAggregate ( dbQuery , new MeanAggregator (2000,new TimeUnits().asMilliseconds())
new TimeQuery(0L,1239867L,”cabin-temperature”)
Interval duration
Unit type
new Database(<data table>,<uid table>, zkQuorum, zkBasePath)
IAggregator Interface

19. SQL based access
SQL Pivotal
HAWQ
Hbase
UDF [
PL-Java ]
Java Read
APIs
RStudio
● Dual Advantages – MPP scaling + underlying
columnar storage
● Security (Kerberos) JVM specific
● PostGreSql package in R (access by data
scientists)
● Scaled out in terms of compute and storage
SQL access for RDBMS based flows
CRAN Project RPostgreSQL: R interface to the PostgreSQL database system
Database interface and PostgreSQL driver for R This package provides a Database Interface
(DBI) compliant driver for R to access PostgreSQL database systems …….

20. SQL based access – User Scenarios
SELECT * FROM getTags('tsdb','tsdb-uid') WHERE gettags LIKE 'CabAttribute%’
SELECT * FROM getTimeSeriesData( 'tsdb','tsdb-uid',CAST('2012-10-01 01:00:00' AS TIMESTAMP ),'temperature5') limit 10
SELECT * FROM getTimeSeriesData( 'tsdb','tsdb-uid',CAST('2012-09-20 01:00:00' AS TIMESTAMP ),CAST('2012-09-21
23:59:59' AS TIMESTAMP ),'temperature5')
SELECT * FROM getTimeSeriesData( 'tsdb','tsdb-uid',CAST('2012-09-20 01:00:00' AS TIMESTAMP ),CAST('2012-09-21
23:59:59' AS TIMESTAMP ),'temperature5','attrib1=*,attrib2=W*")
Title or Job Number | XX Month 201X 20
Time
Range
Tag Input Attributes Return Attributes
Start
Time,
[End
Time]
Tag Only None None
Start
Time,
[End
Time]
Tag +
Attributes
All(*), key=value, key=<regex> All
key=value,
[p=q,x=y,…]
key=value1,key=valu
e2, [p=q,x=y, …]
p,q,x,y,… other
attributes in time
series data point
TimeStamp Tag Attribute
Name
Attribute
Value
2012-10-01
1:00:00
Tempe
rature
5
Attrib1 v1
2012-10-01
1:00:00
Tempe
rature
5
Attrib1 v2
2012-10-01
1:00:00
Tempe
rature
5
Attrib2 W1
2012-10-01
1:00:00
Tempe
arture
5
Attrib2 W2
2012-11-01 Tempe
rature
5
… …

21. Data Models – HBase PoC
• Nature of sensor data from engines - sparse
• Horizontal v/s vertical data model
• Vertical – 1 flight parameter / column (1-2K values)
• Horizontal – Parameters converted to rows, needs transposition
during the ingestion
• Performance on Hbase
• Horizontal model did much better for retrieval
• HAWQ & HAWQ as external table over Hbase was
slower

22. Recap / Summary
• Industrial Data – nature of the business problem
• Industrial Data Lake
• Technical Solution
• Wrap up