1. dbTouch: Analytics at your Fingertips
Stratos Idreos Erietta Liarou
International Conference on Innovative Data Systems Research (CIDR). Asilomar, California; 2013.

2. Big Data Scenario Big, frequent, various forms
not always sure what we are looking
Every two days we create as much data as we did from dawn of humanity to [Eric Schimdt]
4.4 Million of new data scientists will be needed by [IBMbigData]
[DOMO]

3. Problems Modern database systems Big data has worsened the situation
strict and monolithic
difficult to use
even more difficult to explore
Big data has worsened the situation
correct answer  slower response
novice analysts  unusable
interactive analytics  impossible
Database Architectures for Big Data Exploration, Stratos Idreos, Keynote in French Database Conference, October, 2013

4. Motivation
making database systems more usable and interactive

5. Vision

6. dbTouch touch based data exploration redesign database kernels
interactive
easy to use
redesign database kernels
tailored for touch based exploration

7. dbTouch
Front-end

8. Data Objects abstract representation of data
tables  fat rectangles
attributes  columns
enables interaction between user and dbTouch
apply various gestures on data objects
interaction through gestures
slide
zoom in/out
rotate

9. Gestures

10. dbTouch in Action
Touch, Crack and Explore, Stratos Idreos, High Performance Transaction Systems (HPTS), September, 2013

11. dbTouch in Action (Accessing a tuple)
Touch, Crack and Explore, Stratos Idreos, High Performance Transaction Systems (HPTS), September, 2013

12. Schema-less Querying exact schema information less important
data objects convey the schema information at a high level
type of data available for exploration (column)
different attribute types (tuple)
some information available on demand
attribute, table name
foreign key relationships

13. dbTouch
Querying through gesture

14. Slide multiple “single taps”
each single tap returns a specific data (data point or tuple)
equivalent to next operation in traditional database systems
sliding speed and direction act as query parameters

15. Query Processing querying in dbTouch not a monolithic action
not all data are processed
query processing can stop at any time
dbTouch query
user defines query through query action
scan, aggregate etc.
start slide gesture over a column or table

16. Slide Example

17. Challenges Which data tuples exactly do we process with every touch?
How should data be stored and accessed?
What happens when the slide patterns such as speed and directions change?
How to maintain quick response times?

18. dbTouch
From Touch to Tuple Identifiers

19. Object Views exploits the view concept of touch-based OS
views are placeholder for visual objects
view hierarchies
master view
views within master view
gestures in each view treated independently of master view

20. Map(Touch, RowID) Rule of three data object refers to column
o = size of the data object (column/ table)
n = number of total tuples in (column/ table)
t = touch location
id = n*t/o
data object refers to column
use height dimension
data object refers to table
tuple scan: use height dimension
horizontal slide (select specific columns)
use both height (tuple) and width (column) dimension
What happens when we zoom in/out and then slide?

21. dbTouch
Touch Granularity

22. Sampling Through Touch
sliding speed  degree of data exploration
faster slide  less touch location registered
slower slide  more touch location registered
visual object size also controls the exploration
only a limited number of tuples mapped and processed
zoom in  more touch location registered
zoom out  less touch location registered
Fine grained and coarse grained touch granularity

23. dbTouch
Storing and Accessing Data

24. Gesture Evolution varying gestures gestures are user dependent
tap, slide, zoom in/out, rotate
gestures are user dependent
may change mid gesture
pause, change sliding direction/speed
underlying storage and access methods
need to adjust accordingly

25. Physical Layout underlying storage model id = n*t/o
row store, column store or a hybrid format
fixed-width fields per attribute
faster Map(Touch, RowID)
underlying storage layout: matrix
one matrix per data object
id = n*t/o

26. Optimizations Sample based storage Pre-fetch data
query processing in dbTouch  processing a sample of data
many random disk seeks
accessing data directly from base table
many unnecessary data loaded
solution: store various different samples of base data
depends on object size and gesture speed
but so many different combination !!!
Pre-fetch data
when a slide pauses or slows down
extrapolate gesture progression (speed and direction)
fetch expected entries

27. Optimizations (cont.) Caching data Indexing
used when user re-examines a data
continually observe gesture progression (speed and direction)
used to generate new sample copies
Indexing
maintain separate index for each sample: sliding  index scan
switching between different indexes - challenging

28. dbTouch
Interactive Summaries

29. Aggregate Queries select the aggregate function, then slide
register position p
read tuples from (idp – k) to (idp + k)
allows user to inspect more data with single “touch”
users can observe
patterns of a small group of data
difference in properties of various groups

30. dbTouch
Schema and Storage Layout Gestures

31. Playing with the Schema
change schema: create table
drag columns at random, drop in table placeholder
change layout
rotate (tablet): row store  column store
project attributes to individual arrays
drag column(s) out of the table
treat as array(smaller table)

32. dbTouch
Complex Queries

33. Complex Queries and Optimizations
enable the query action: aggregate, where clause, join etc.
joins are difficult
use entire input  delay
cannot perform hash join
doesn’t know which data to use for the hash table
cache hash tables across various sample copies
data to be processed unknown
makes optimization difficult
adaptive optimization interleaving with query execution
take optimization decisions on the fly
maintain good response times

34. dbTouch
Implementation and Experiments

35. dbTouch Prototype
small prototype database kernel implemented on top of IOs
dbTouch system flow same for all touch actions
recognize touch
recognize gesture
map touch to data
execute

36. Experiments

37. dbTouch
Future Work

38. Challenges and Opportunities
Approximate query processing in combination with dbTouch
provide results in expected response time
continually refine results
Include charts besides raw data visualization
Remote processing
touch device is the interface to a cloud facility
use local data for interactive response
request remote data for further exploration
Alternative interfaces
other forms of input: motion, speech etc.

39. Touch, Crack and Explore, Stratos Idreos, High Performance Transaction Systems (HPTS), September, 2013

40. Comments
Does it scale?
higher granularity  interesting pattern skipped
user study
complex queries
multiple aggregates grouped by some column
experiments
interactive response?

41. Thank you