SLIDE 1
FROM RETROSPECTIVE TO CONTINUOUS DEEP ANALYTICS Seif Haridi KTH - - PowerPoint PPT Presentation
FROM RETROSPECTIVE TO CONTINUOUS DEEP ANALYTICS Seif Haridi KTH - - PowerPoint PPT Presentation
FROM RETROSPECTIVE TO CONTINUOUS DEEP ANALYTICS Seif Haridi KTH SICS Why most Data Analysis today is Retrospective From OLAP Databases Data Cube LOCATION TIME To Deep Analysis on Batches Past Data Sets Map Map Map Map Map Map
SLIDE 2
SLIDE 3
From OLAP Databases
TIME LOCATION Data Cube
SLIDE 4
To Deep Analysis on Batches
Map Reduce Map Reduce Map Reduce Map Reduce Map Reduce Map Reduce
Past Data Sets Past Models
SLIDE 5
And Machine Learning Pipelines
Yesterday’s Dataset Feature Extraction Model Training
SLIDE 6
Today’s Retrospective Data Processing
STORE LOAD PROCESS STORE Data Knowledge
SLIDE 7
Today’s Retrospective Data Processing Takes Long to Extract Knowledge
Data Knowledge time
SLIDE 8
Models and patterns on older data are often irrelevant today
personalized ad winter jacket discount!
SLIDE 9
Critical Decisions demand Continuous Analysis
SLIDE 10
We propose a continuous processing architecture
STORE LOAD PROCESS STORE time Data Knowledge
SLIDE 11
PROCESS
We propose a continuous processing architecture
time Data Knowledge
SLIDE 12
PROCESS
We propose a deep processing architecture
arbitrarily iterative computation
SLIDE 13
PROCESS TPUs
We propose a fast processing architecture
SLIDE 14
for Dynamic Graph analysis
virus outbreaks social network trends
SLIDE 15
for Online Machine Learning
feature learning tensor programming
SLIDE 16
and Relational Data Streaming
dynamic tables
σθ σθ σθ σθ π π
SLIDE 17
Unified and Optimised
- n a Common Representation
Intermediate Representation Dynamic Graphs Online ML Relational Streams
SLIDE 18
A Runtime Designed for Continuous Deep Analysis
Distributed Runtime Intermediate Representation
metrics config constraints
Self‐Reconfiguration
SLIDE 19