Guest lecture 2013-08-27 at General Assembly in SF for the Data Science program taught by Jacob Bollinger and Thomson Nguyen https://generalassemb.ly/education/data-science/san-francisco
Many thanks to Thomson, Jacob, and the participants in the course. Excellent Q&A!
Received a bottle o' Cardhu (my fave Scotch) in payment for lecture, and since it's Burning Man Week, the city was emptied so we had enough to share with the class :)
Evidence:
https://plus.google.com/u/0/110794698656267747127/posts/GvjhhQ99CTs
Data Center Computing for Data Science: an evolution of machines, middleware, math, and Mesos
1. General Assembly SF, 2013-08-27:
“Data Center Computing for Data Science:
an evolution of machines, middleware, math, and Mesos”
Learnings generalized from trends in Data Science:
a 30-year retrospective on Machine Learning,
a 10-year summary of Leading Data ScienceTeams,
and a 2-year survey of Enterprise Use Cases
Paco Nathan @pacoid
Chief Scientist, Mesosphere
1Saturday, 31 August 13
2. Learnings generalized from trends in Data Science:
1. the practice of leading data science teams
2. strategies for leveraging data at scale
3. machine learning and optimization
4. large-scale data workflows
5. an evolution of cluster computing
GA/SF, 2013-08-27
2Saturday, 31 August 13
3. employing a mode of thought which includes both logical and analytical reasoning:
evaluating the whole of a problem, as well as its component parts; attempting
to assess the effects of changing one or more variables
this approach attempts to understand not just problems and solutions,
but also the processes involved and their variances
particularly valuable in Big Data work when combined with hands-on experience in
physics – roughly 50% of my peers come from physics or physical engineering…
programmers typically don’t think this way…
however, both systems engineers and data scientists must
Process Variation Data Tools
Statistical Thinking
3Saturday, 31 August 13
4. Modeling
back in the day, we worked with practices based on
data modeling
1. sample the data
2. fit the sample to a known distribution
3. ignore the rest of the data
4. infer, based on that fitted distribution
that served well with ONE computer, ONE analyst,
ONE model… just throw away annoying “extra” data
circa late 1990s: machine data, aggregation, clusters, etc.
algorithmic modeling displaced the prior practices
of data modeling
because the data won’t fit on one computer anymore
4Saturday, 31 August 13
5. Two Cultures
“A new research community using these tools sprang up.Their goal
was predictive accuracy.The community consisted of young computer
scientists, physicists and engineers plus a few aging statisticians.
They began using the new tools in working on complex prediction
problems where it was obvious that data models were not applicable:
speech recognition, image recognition, nonlinear time series prediction,
handwriting recognition, prediction in financial markets.”
Statistical Modeling: TheTwo Cultures
Leo Breiman, 2001
bit.ly/eUTh9L
chronicled a sea change from data modeling (silos, manual
process) to the rising use of algorithmic modeling (machine
data for automation/optimization) which led in turn to the
practice of leveraging inter-disciplinary teams
5Saturday, 31 August 13
6. approximately 80% of the costs for data-related projects
gets spent on data preparation – mostly on cleaning up
data quality issues: ETL, log files, etc., generally by socializing
the problem
unfortunately, data-related budgets tend to go into
frameworks that can only be used after clean up
most valuable skills:
‣ learn to use programmable tools that prepare data
‣ learn to understand the audience and their priorities
‣ learn to socialize the problems, knocking down silos
‣ learn to generate compelling data visualizations
‣ learn to estimate the confidence for reported results
‣ learn to automate work, making process repeatable
What is needed most?
UniqueRegistration
aunchedgameslobby
NUI:TutorialMode
BirthdayMessage
hatPublicRoomvoice
unchedheyzapgame
Test:testsuitestarted
CreateNewPet
rted:client,community
NUI:MovieMode
BuyanItem:web
PutonClothing
paceremaining:512M
aseCartPageStep2
FeedPet
PlayPet
ChatNow
EditPanel
anelFlipProductOver
AddFriend
Open3DWindow
ChangeSeat
TypeaBubble
VisitOwnHomepage
TakeaSnapshot
NUI:BuyCreditsMode
NUI:MyProfileClicked
sspaceremaining:1G
LeaveaMessage
NUI:ChatMode
NUI:FriendsMode
dv
WebsiteLogin
AddBuddy
NUI:PublicRoomMode
NUI:MyRoomMode
anelRemoveProduct
yPanelApplyProduct
NUI:DressUpMode
UniqueRegistration
Launchedgameslobby
NUI:TutorialMode
BirthdayMessage
ChatPublicRoomvoice
Launchedheyzapgame
ConnectivityTest:testsuitestarted
CreateNewPet
MovieViewStarted:client,community
NUI:MovieMode
BuyanItem:web
PutonClothing
Addressspaceremaining:512M
CustomerMadePurchaseCartPageStep2
FeedPet
PlayPet
ChatNow
EditPanel
ClientInventoryPanelFlipProductOver
AddFriend
Open3DWindow
ChangeSeat
TypeaBubble
VisitOwnHomepage
TakeaSnapshot
NUI:BuyCreditsMode
NUI:MyProfileClicked
Addressspaceremaining:1G
LeaveaMessage
NUI:ChatMode
NUI:FriendsMode
dv
WebsiteLogin
AddBuddy
NUI:PublicRoomMode
NUI:MyRoomMode
ClientInventoryPanelRemoveProduct
ClientInventoryPanelApplyProduct
NUI:DressUpMode
6Saturday, 31 August 13
7. apps
discovery
modeling
integration
systems
help people ask the
right questions
allow automation to
place informed bets
deliver data products
at scale to LOB end uses
build smarts into
product features
keep infrastructure
running, cost-effective
Team Process = Needs
analysts
engineers
inter-disciplinary
leadership
7Saturday, 31 August 13
8. business process,
stakeholder
data prep, discovery,
modeling, etc.
software engineering,
automation
systems engineering,
availability
data
science
Data
Scientist
App Dev
Ops
Domain
Expert
introduced
capability
Team Composition = Roles
leverage non-traditional
pairing among roles, to
complement skills and
tear down silos
8Saturday, 31 August 13
10. Alternatively, Data Roles × Skill Sets
Harlan Harris, et al.
datacommunitydc.org/blog/wp-content/uploads/
2012/08/SkillsSelfIDMosaic-edit-500px.png
Analyzing the Analyzers
Harlan Harris, Sean Murphy,
Marck Vaisman
O’Reilly, 2013
amazon.com/dp/B00DBHTE56
10Saturday, 31 August 13
11. Learning Curves
difficulties in the commercial use of distributed systems
often get represented as issues of managing complexity
much of the risk in managing a data science team is about
budgeting for learning curve: some orgs practice a kind of
engineering “conservatism”, with highly structured process
and strictly codified practices – people learn a few things
well, then avoid having to struggle with learning many new
things perpetually…
that anti-pattern leads to big teams, low ROI
scale➞
complexity➞
ultimately, the challenge is about
managing learning curves within
a social context
11Saturday, 31 August 13
12. Learnings generalized from trends in Data Science:
1. the practice of leading data science teams
2. strategies for leveraging data at scale
3. machine learning and optimization
4. large-scale data workflows
5. an evolution of cluster computing
GA/SF, 2013-08-27
12Saturday, 31 August 13
13. Business Disruption through Data
Geoffrey Moore
Mohr DavidowVentures, author CrossingThe Chasm
@Hadoop Summit, 2012:
what Amazon did to the retail sector… has put the
entire Global 1000 on notice over the next decade…
data as the major force… mostly through apps –
verticals, leveraging domain expertise
Michael Stonebraker
INGRES, PostgreSQL,Vertica,VoltDB, Paradigm4, etc.
@XLDB, 2012:
complex analytics workloads are now displacing SQL
as the basis for Enterprise apps
13Saturday, 31 August 13
14. Data Categories
Three broad categories of data
Curt Monash, 2010
dbms2.com/2010/01/17/three-broad-categories-of-data
• Human/Tabular data – human-generated data which fits into tables/arrays
• Human/Nontabular data – all other data generated by humans
• Machine-Generated data
let’s now add other useful distinctions:
• Open Data
• Curated Metadata
• A/D conversion for sensors (IoT)
14Saturday, 31 August 13
15. Open Data notes
successful apps incorporate three components:
• Big Data (consumer interest, personalization)
• Open Data (monetizing public data)
• Curated Metadata
most of the largest Cascading deployments leverage some
Open Data components: Climate Corp, Factual, Nokia, etc.
consider buildingeye.com, aggregate building permits:
• pricing data for home owners looking to remodel
• sales data for contractors
• imagine joining data with building inspection history,
for better insights about properties for sale…
research notes about
Open Data use cases:
goo.gl/cd995T
15Saturday, 31 August 13
16. Trends in Public Administration
late 1880s – late 1920s (Woodrow Wilson)
as hierarchy, bureaucracy → only for the most educated, elite
late 1920s – late 1930s
as a business, relying on “Scientific Method”, gov as a process
late 1930s – late 1940s (Robert Dale)
relationships, behavioral-based → policy not separate from politics
late 1940s – 1980s
yet another form of management → less “command and control”
1980s – 1990s (David Osborne,Ted Gaebler)
New Public Management → service efficiency, more private sector
1990s – present (Janet & Robert Denhardt)
Digital Age → transparency, citizen-based “debugging”, bankruptcies
Adapted from:
The Roles,Actors, and Norms Necessary to
Institutionalize Sustainable Collaborative Governance
Peter Pirnejad
USC Price School of Policy
2013-05-02
Drivers, circa 2013
• governments have run out of money,
cannot increase staff and services
• better data infra at scale (cloud, OSS, etc.)
• machine learning techniques to monetize
• viable ecosystem for data products,APIs
• mobile devices enabling use cases
16Saturday, 31 August 13
17. Open Data ecosystem
municipal
departments
publishing
platforms
aggregators
data product
vendors
end use
cases
e.g., Palo Alto, Chicago, DC, etc.
e.g., Junar, Socrata, etc.
e.g., OpenStreetMap,WalkScore, etc.
e.g., Factual, Marinexplore, etc.
e.g., Facebook, Climate, etc.
Data feeds structured for
public private partnerships
17Saturday, 31 August 13
18. Open Data ecosystem – caveats for agencies
municipal
departments
publishing
platforms
aggregators
data product
vendors
end use
cases
e.g., Palo Alto, Chicago, DC, etc.
e.g., Junar, Socrata, etc.
e.g., OpenStreetMap,WalkScore, etc.
e.g., Factual, Marinexplore, etc.
e.g., Facebook, Climate, etc.
Required Focus
• respond to viable use cases
• not budgeting hackathons
18Saturday, 31 August 13
19. Open Data ecosystem – caveats for publishers
municipal
departments
publishing
platforms
aggregators
data product
vendors
end use
cases
e.g., Palo Alto, Chicago, DC, etc.
e.g., Junar, Socrata, etc.
e.g., OpenStreetMap,WalkScore, etc.
e.g., Factual, Marinexplore, etc.
e.g., Facebook, Climate, etc.
Required Focus
• surface the metadata
• curate, allowing for joins/aggregation
• not scans as PDFs
19Saturday, 31 August 13
20. Open Data ecosystem – caveats for aggregators
municipal
departments
publishing
platforms
aggregators
data product
vendors
end use
cases
e.g., Palo Alto, Chicago, DC, etc.
e.g., Junar, Socrata, etc.
e.g., OpenStreetMap,WalkScore, etc.
e.g., Factual, Marinexplore, etc.
e.g., Facebook, Climate, etc.
Required Focus
• make APIs consumable by automation
• allow for probabilistic usage
• not OSS licensing for data
20Saturday, 31 August 13
21. Open Data ecosystem – caveats for data vendors
municipal
departments
publishing
platforms
aggregators
data product
vendors
end use
cases
e.g., Palo Alto, Chicago, DC, etc.
e.g., Junar, Socrata, etc.
e.g., OpenStreetMap,WalkScore, etc.
e.g., Factual, Marinexplore, etc.
e.g., Facebook, Climate, etc.
Required Focus
• supply actionable data
• track data provenance carefully
• provide feedback upstream,
i.e., cleaned data at source
• focus on core verticals
21Saturday, 31 August 13
22. Open Data ecosystem – caveats for end uses
municipal
departments
publishing
platforms
aggregators
data product
vendors
end use
cases
e.g., Palo Alto, Chicago, DC, etc.
e.g., Junar, Socrata, etc.
e.g., OpenStreetMap,WalkScore, etc.
e.g., Factual, Marinexplore, etc.
e.g., Facebook, Climate, etc.
Required Focus
• address consumer needs
• identify community benefits
of the data
22Saturday, 31 August 13
23. algorithmic modeling
+ machine data (Big Data)
+ curation, metadata
+ Open Data
data products, as feedback into automation
evolution of feedback loops
less about “bigness”, more about complexity
internet of things
+ A/D conversion
+ more complex analytics
accelerated evolution, additional feedback loops
orders of magnitude higher data rates
Recipes for Success
source: National Geographic
“A kind of Cambrian explosion”
source: National Geographic
23Saturday, 31 August 13
24. Trendlines
Big Data? we’re just getting started:
• ~12 exabytes/day, jet turbines on commercial flights
• Google self-driving cars, ~1 Gb/s per vehicle
• National Instruments initiative: Big Analog Data™
• 1m resolution satellites skyboximaging.com
• open resource monitoring reddmetrics.com
• Sensing XChallenge nokiasensingxchallenge.org
consider the implications of Jawbone, Nike, etc.,
plus the effects of Google Glass…
technologyreview.com/...
24Saturday, 31 August 13
26. Learnings generalized from trends in Data Science:
1. the practice of leading data science teams
2. strategies for leveraging data at scale
3. machine learning and optimization
4. large-scale data workflows
5. an evolution of cluster computing
GA/SF, 2013-08-27
26Saturday, 31 August 13
27. in general, apps alternate between learning patterns/rules
and retrieving similar things…
machine learning – scalable, arguably quite ad-hoc,
generally “black box” solutions, enabling you to make billion
dollar mistakes, with oh so much commercial emphasis
(i.e. the “heavy lifting”)
statistics – rigorous, much slower to evolve, confidence
and rationale become transparent, preventing you from
making billion dollar mistakes, any good commercial project
has ample stats work used in QA
(i.e.,“CYA, cover your analysis”)
once Big Data projects get beyond merely digesting
log files, optimization will likely become the next
overused buzzword :)
Learning Theory
27Saturday, 31 August 13
28. Generalizations about Machine Learning…
great introduction to ML, plus a proposed categorization
for comparing different machine learning approaches:
A Few UsefulThings to Know about Machine Learning
Pedro Domingos, U Washington
homes.cs.washington.edu/~pedrod/papers/cacm12.pdf
toward a categorization for Machine Learning algorithms:
• representation: classifier must be represented in some
formal language that computers can handle (algorithms, data
structures, etc.)
• evaluation: evaluation function (objective function, scoring
function) is needed to distinguish good classifiers from bad
ones
• optimization: method to search among the classifiers in
the language for the highest-scoring one
28Saturday, 31 August 13
29. Something to consider about Algorithms…
many algorithm libraries used today are based on implementations
back when people used DO loops in FORTRAN, 30+ years ago
MapReduce is Good Enough?
Jimmy Lin, U Maryland
umiacs.umd.edu/~jimmylin/publications/Lin_BigData2013.pdf
astrophysics and genomics are light years ahead of e-commerce in
terms of data rates and sophisticated algorithms work – as Breiman
suggested in 2001 – may take a few years to percolate into industry
other game-changers:
• streaming algorithms, sketches, probabilistic data structures
• significant “Big O” complexity reduction (e.g., skytree.net)
• better architectures and topologies (e.g., GPUs and CUDA)
• partial aggregates – parallelizing workflows
29Saturday, 31 August 13
30. Make It Sparse…
also, take a moment to check this out…
(and related work on sparse Cholesky, etc.)
QR factorization of a “tall-and-skinny” matrix
• used to solve many data problems at scale,
e.g., PCA, SVD, etc.
• numerically stable with efficient implementation
on large-scale Hadoop clusters
suppose that you have a sparse matrix of customer
interactions where there are 100MM customers,
with a limited set of outcomes…
cs.purdue.edu/homes/dgleich
stanford.edu/~arbenson
github.com/ccsevers/scalding-linalg
David Gleich, slideshare.net/dgleich
30Saturday, 31 August 13
31. Sparse Matrix Collection
for those times when you really, really need
a wide variety of sparse matrix examples…
University of Florida Sparse Matrix Collection
cise.ufl.edu/research/sparse/matrices/
Tim Davis, U Florida
cise.ufl.edu/~davis/welcome.html
Yifan Hu, AT&T Research
www2.research.att.com/~yifanhu/
31Saturday, 31 August 13
32. A Winning Approach…
consider that if you know priors about a system, then
you may be able to leverage low dimensional structure
within high dimensional data… what impact does that
have on sampling rates?
1. real-world data
2. graph theory for representation
3. sparse matrix factorization for production work
4. cost-effective parallel processing
for machine learning app at scale
32Saturday, 31 August 13
33. Just Enough Mathematics?
having a solid background in statistics becomes vital,
because it provides formalisms for what we’re trying
to accomplish at scale
along with that, some areas of math help – regardless
of the “calculus threshold” invoked at many universities…
linear algebra e.g., calculating algorithms for large-scale apps efficiently
graph theory e.g., representation of problems in a calculable language
abstract algebra e.g., probabilistic data structures in streaming analytics
topology e.g., determining the underlying structure of the data
operations research e.g., techniques for optimization … in other words, ROI
33Saturday, 31 August 13
34. ADMM: a general approach for optimizing learners
Distributed Optimization and Statistical Learning
via the Alternating Direction Method of Multipliers
Stephen Boyd, Neal Parikh, et al., Stanford
stanford.edu/~boyd/papers/admm_distr_stats.html
“Throughout, the focus is on applications rather than theory, and a main goal is
to provide the reader with a kind of ‘toolbox’ that can be applied in many situations
to derive and implement a distributed algorithm of practical use.Though the focus
here is on parallelism, the algorithm can also be used serially, and it is interesting
to note that with no tuning, ADMM can be competitive with the best known
methods for some problems.”
“While we have emphasized applications that can be concisely explained, the
algorithm would also be a natural fit for more complicated problems in areas
like graphical models. In addition, though our focus is on statistical learning
problems, the algorithm is readily applicable in many other cases, such as in
engineering design, multi-period portfolio optimization, time series analysis,
network flow, or scheduling.”
34Saturday, 31 August 13
35. Learnings generalized from trends in Data Science:
1. the practice of leading data science teams
2. strategies for leveraging data at scale
3. machine learning and optimization
4. large-scale data workflows
5. an evolution of cluster computing
GA/SF, 2013-08-27
35Saturday, 31 August 13
36. Enterprise Data Workflows
middleware for Big Data applications is evolving,
with commercial examples that include:
Cascading, Lingual, Pattern, etc.
Concurrent
ParAccel Big Data Analytics Platform
Actian
Anaconda supporting IPython Notebook, Pandas,Augustus, etc.
Continuum Analytics
ETL
data
prep
predictive
model
data
sources
end
uses
36Saturday, 31 August 13
37. Anatomy of an Enterprise app
definition of a typical Enterprise workflow which crosses through
multiple departments, languages, and technologies…
ETL
data
prep
predictive
model
data
sources
end
uses
ANSI SQL for ETL
37Saturday, 31 August 13
38. Anatomy of an Enterprise app
definition of a typical Enterprise workflow which crosses through
multiple departments, languages, and technologies…
ETL
data
prep
predictive
model
data
sources
end
usesJ2EE for business logic
38Saturday, 31 August 13
39. Anatomy of an Enterprise app
definition of a typical Enterprise workflow which crosses through
multiple departments, languages, and technologies…
ETL
data
prep
predictive
model
data
sources
end
uses
SAS for predictive models
39Saturday, 31 August 13
40. Anatomy of an Enterprise app
definition of a typical Enterprise workflow which crosses through
multiple departments, languages, and technologies…
ETL
data
prep
predictive
model
data
sources
end
uses
SAS for predictive modelsANSI SQL for ETL most of the licensing costs…
40Saturday, 31 August 13
41. Anatomy of an Enterprise app
definition of a typical Enterprise workflow which crosses through
multiple departments, languages, and technologies…
ETL
data
prep
predictive
model
data
sources
end
usesJ2EE for business logic
most of the project costs…
41Saturday, 31 August 13
42. ETL
data
prep
predictive
model
data
sources
end
uses
Lingual:
DW → ANSI SQL
Pattern:
SAS, R, etc. → PMML
business logic in Java,
Clojure, Scala, etc.
sink taps for
Memcached, HBase,
MongoDB, etc.
source taps for
Cassandra, JDBC,
Splunk, etc.
Anatomy of an Enterprise app
Cascading allows multiple departments to combine their workflow components
into an integrated app – one among many, typically – based on 100% open source
a compiler sees it all…
one connected DAG:
• optimization
• troubleshooting
• exception handling
• notifications
cascading.org
42Saturday, 31 August 13
43. a compiler sees it all…
ETL
data
prep
predictive
model
data
sources
end
uses
Lingual:
DW → ANSI SQL
Pattern:
SAS, R, etc. → PMML
business logic in Java,
Clojure, Scala, etc.
sink taps for
Memcached, HBase,
MongoDB, etc.
source taps for
Cassandra, JDBC,
Splunk, etc.
Anatomy of an Enterprise app
Cascading allows multiple departments to combine their workflow components
into an integrated app – one among many, typically – based on 100% open source
FlowDef flowDef = FlowDef.flowDef()
.setName( "etl" )
.addSource( "example.employee", emplTap )
.addSource( "example.sales", salesTap )
.addSink( "results", resultsTap );
SQLPlanner sqlPlanner = new SQLPlanner()
.setSql( sqlStatement );
flowDef.addAssemblyPlanner( sqlPlanner );
cascading.org
43Saturday, 31 August 13
44. a compiler sees it all…
ETL
data
prep
predictive
model
data
sources
end
uses
Lingual:
DW → ANSI SQL
Pattern:
SAS, R, etc. → PMML
business logic in Java,
Clojure, Scala, etc.
sink taps for
Memcached, HBase,
MongoDB, etc.
source taps for
Cassandra, JDBC,
Splunk, etc.
Anatomy of an Enterprise app
Cascading allows multiple departments to combine their workflow components
into an integrated app – one among many, typically – based on 100% open source
FlowDef flowDef = FlowDef.flowDef()
.setName( "classifier" )
.addSource( "input", inputTap )
.addSink( "classify", classifyTap );
PMMLPlanner pmmlPlanner = new PMMLPlanner()
.setPMMLInput( new File( pmmlModel ) )
.retainOnlyActiveIncomingFields();
flowDef.addAssemblyPlanner( pmmlPlanner );
44Saturday, 31 August 13
45. Cascading – functional programming
Key insight: MapReduce is based on functional programming
– back to LISP in 1970s. Apache Hadoop use cases are
mostly about data pipelines, which are functional in nature.
to ease staffing problems as “Main Street” Enterprise firms
began to embrace Hadoop, Cascading was introduced
in late 2007, as a new Java API to implement functional
programming for large-scale data workflows:
• leverages JVM and Java-based tools without any
need to create new languages
• allows programmers who have J2EE expertise
to leverage the economics of Hadoop clusters
Edgar Codd alluded to this (DSLs for structuring data)
in his original paper about relational model
45Saturday, 31 August 13
46. Cascading – functional programming
• Twitter, eBay, LinkedIn, Nokia, YieldBot, uSwitch, etc.,
have invested in open source projects atop Cascading –
used for their large-scale production deployments
• new case studies for Cascading apps are mostly based on
domain-specific languages (DSLs) in JVM languages which
emphasize functional programming:
Cascalog in Clojure (2010)
Scalding in Scala (2012)
github.com/nathanmarz/cascalog/wiki
github.com/twitter/scalding/wiki
Why Adopting the Declarative Programming PracticesWill ImproveYour Return fromTechnology
Dan Woods, 2013-04-17 Forbes
forbes.com/sites/danwoods/2013/04/17/why-adopting-the-declarative-programming-
practices-will-improve-your-return-from-technology/
46Saturday, 31 August 13
47. Workflow Abstraction – pattern language
Cascading uses a “plumbing” metaphor in Java
to define workflows out of familiar elements:
Pipes, Taps, Tuple Flows, Filters, Joins, Traps, etc.
Scrub
token
Document
Collection
Tokenize
Word
Count
GroupBy
token
Count
Stop Word
List
Regex
token
HashJoin
Left
RHS
M
R
data is represented as flows of tuples
operations in the flows bring functional
programming aspects into Java
A Pattern Language
Christopher Alexander, et al.
amazon.com/dp/0195019199
47Saturday, 31 August 13
48. Workflow Abstraction – business process
following the essence of literate programming, Cascading
workflows provide statements of business process
this recalls a sense of business process management
for Enterprise apps (think BPM/BPEL for Big Data)
Cascading creates a separation of concerns between
business process and implementation details (Hadoop, etc.)
this is especially apparent in large-scale Cascalog apps:
“Specify what you require, not how to achieve it.”
by virtue of the pattern language, the flow planner then
determines how to translate business process into efficient,
parallel jobs at scale
48Saturday, 31 August 13
49. void map (String doc_id, String text):
for each word w in segment(text):
emit(w, "1");
void reduce (String word, Iterator group):
int count = 0;
for each pc in group:
count += Int(pc);
emit(word, String(count));
The Ubiquitous Word Count
Definition:
this simple program provides an excellent test case
for parallel processing:
• requires a minimal amount of code
• demonstrates use of both symbolic and numeric values
• shows a dependency graph of tuples as an abstraction
• is not many steps away from useful search indexing
• serves as a “HelloWorld” for Hadoop apps
a distributed computing framework that runsWord Count
efficiently in parallel at scale can handle much larger
and more interesting compute problems
count how often each word appears
in a collection of text documents
49Saturday, 31 August 13
53. (ns impatient.core
(:use [cascalog.api]
[cascalog.more-taps :only (hfs-delimited)])
(:require [clojure.string :as s]
[cascalog.ops :as c])
(:gen-class))
(defmapcatop split [line]
"reads in a line of string and splits it by regex"
(s/split line #"[[](),.)s]+"))
(defn -main [in out & args]
(?<- (hfs-delimited out)
[?word ?count]
((hfs-delimited in :skip-header? true) _ ?line)
(split ?line :> ?word)
(c/count ?count)))
; Paul Lam
; github.com/Quantisan/Impatient
WordCount – Cascalog / Clojure
Document
Collection
Word
Count
Tokenize
GroupBy
token Count
R
M
53Saturday, 31 August 13
54. github.com/nathanmarz/cascalog/wiki
• implements Datalog in Clojure, with predicates backed
by Cascading – for a highly declarative language
• run ad-hoc queries from the Clojure REPL –
approx. 10:1 code reduction compared with SQL
• composable subqueries, used for test-driven development
(TDD) practices at scale
• Leiningen build: simple, no surprises, in Clojure itself
• more new deployments than other Cascading DSLs –
Climate Corp is largest use case: 90% Clojure/Cascalog
• has a learning curve, limited number of Clojure developers
• aggregators are the magic, and those take effort to learn
WordCount – Cascalog / Clojure
Document
Collection
Word
Count
Tokenize
GroupBy
token Count
R
M
54Saturday, 31 August 13
55. import com.twitter.scalding._
class WordCount(args : Args) extends Job(args) {
Tsv(args("doc"),
('doc_id, 'text),
skipHeader = true)
.read
.flatMap('text -> 'token) {
text : String => text.split("[ [](),.]")
}
.groupBy('token) { _.size('count) }
.write(Tsv(args("wc"), writeHeader = true))
}
WordCount – Scalding / Scala
Document
Collection
Word
Count
Tokenize
GroupBy
token Count
R
M
55Saturday, 31 August 13
56. github.com/twitter/scalding/wiki
• extends the Scala collections API so that distributed lists
become “pipes” backed by Cascading
• code is compact, easy to understand
• nearly 1:1 between elements of conceptual flow diagram
and function calls
• extensive libraries are available for linear algebra, abstract
algebra, machine learning – e.g., Matrix API, Algebird, etc.
• significant investments by Twitter, Etsy, eBay, etc.
• great for data services at scale
• less learning curve than Cascalog
WordCount – Scalding / Scala
Document
Collection
Word
Count
Tokenize
GroupBy
token Count
R
M
56Saturday, 31 August 13
57. CREATE TABLE text_docs (line STRING);
LOAD DATA LOCAL INPATH 'data/rain.txt'
OVERWRITE INTO TABLE text_docs
;
SELECT
word, COUNT(*)
FROM
(SELECT
split(line, 't')[1] AS text
FROM text_docs
) t
LATERAL VIEW explode(split(text, '[ ,.()]')) lTable AS
word
GROUP BY word
;
WordCount – Apache Hive
Document
Collection
Word
Count
Tokenize
GroupBy
token Count
R
M
57Saturday, 31 August 13
58. WordCount – Apache Hive
Document
Collection
Word
Count
Tokenize
GroupBy
token Count
R
M
hive.apache.org
pro:
‣ most popular abstraction atop Apache Hadoop
‣ SQL-like language is syntactically familiar to most analysts
‣ simple to load large-scale unstructured data and run ad-hoc queries
con:
‣ not a relational engine, many surprises at scale
‣ difficult to represent complex workflows, ML algorithms, etc.
‣ one poorly-trained analyst can bottleneck an entire cluster
‣ app-level integration requires other coding, outside of script language
‣ logical planner mixed with physical planner; cannot collect app stats
‣ non-deterministic exec: number of maps+reduces may change unexpectedly
‣ business logic must cross multiple language boundaries: difficult to
troubleshoot, optimize, audit, handle exceptions, set notifications, etc.
58Saturday, 31 August 13
59. docPipe = LOAD '$docPath' USING PigStorage('t', 'tagsource')
AS (doc_id, text);
docPipe = FILTER docPipe BY doc_id != 'doc_id';
-- specify regex to split "document" text lines into token stream
tokenPipe = FOREACH docPipe
GENERATE doc_id, FLATTEN(TOKENIZE(text, ' [](),.')) AS token;
tokenPipe = FILTER tokenPipe BY token MATCHES 'w.*';
-- determine the word counts
tokenGroups = GROUP tokenPipe BY token;
wcPipe = FOREACH tokenGroups
GENERATE group AS token, COUNT(tokenPipe) AS count;
-- output
STORE wcPipe INTO '$wcPath' USING PigStorage('t', 'tagsource');
EXPLAIN -out dot/wc_pig.dot -dot wcPipe;
WordCount – Apache Pig
Document
Collection
Word
Count
Tokenize
GroupBy
token Count
R
M
59Saturday, 31 August 13
60. WordCount – Apache Pig
Document
Collection
Word
Count
Tokenize
GroupBy
token Count
R
M
pig.apache.org
pro:
‣ easy to learn data manipulation language (DML)
‣ interactive prompt (Grunt) makes it simple to prototype apps
‣ extensibility through UDFs
con:
‣ not a full programming language; must extend via UDFs outside of language
‣ app-level integration requires other coding, outside of script language
‣ simple problems are simple to do; hard problems become quite complex
‣ difficult to parameterize scripts externally; must rewrite to change taps!
‣ logical planner mixed with physical planner; cannot collect app stats
‣ non-deterministic exec: number of maps+reduces may changes unexpectedly
‣ business logic must cross multiple language boundaries: difficult to
troubleshoot, optimize, audit, handle exceptions, set notifications, etc.
60Saturday, 31 August 13
61. Two Avenues to the App Layer…
scale ➞
complexity➞
Enterprise: must contend with
complexity at scale everyday…
incumbents extend current practices and
infrastructure investments – using J2EE,
ANSI SQL, SAS, etc. – to migrate
workflows onto Apache Hadoop while
leveraging existing staff
Start-ups: crave complexity and
scale to become viable…
new ventures move into Enterprise space
to compete using relatively lean staff,
while leveraging sophisticated engineering
practices, e.g., Cascalog and Scalding
61Saturday, 31 August 13
62. Learnings generalized from trends in Data Science:
1. the practice of leading data science teams
2. strategies for leveraging data at scale
3. machine learning and optimization
4. large-scale data workflows
5. an evolution of cluster computing
GA/SF, 2013-08-27
62Saturday, 31 August 13
63. Q3 1997: inflection point
four independent teams were working toward horizontal
scale-out of workflows based on commodity hardware
this effort prepared the way for huge Internet successes
in the 1997 holiday season… AMZN, EBAY, Inktomi
(YHOO Search), then GOOG
MapReduce and the Apache Hadoop open source stack
emerged from this period
63Saturday, 31 August 13
64. RDBMS
Stakeholder
SQL Query
result sets
Excel pivot tables
PowerPoint slide decks
Web App
Customers
transactions
Product
strategy
Engineering
requirements
BI
Analysts
optimized
code
Circa 1996: pre- inflection point
64Saturday, 31 August 13
65. RDBMS
Stakeholder
SQL Query
result sets
Excel pivot tables
PowerPoint slide decks
Web App
Customers
transactions
Product
strategy
Engineering
requirements
BI
Analysts
optimized
code
Circa 1996: pre- inflection point
“throw it over the wall”
65Saturday, 31 August 13
66. RDBMS
SQL Query
result sets
recommenders
+
classifiers
Web Apps
customer
transactions
Algorithmic
Modeling
Logs
event
history
aggregation
dashboards
Product
Engineering
UX
Stakeholder Customers
DW ETL
Middleware
servletsmodels
Circa 2001: post- big ecommerce successes
66Saturday, 31 August 13
67. RDBMS
SQL Query
result sets
recommenders
+
classifiers
Web Apps
customer
transactions
Algorithmic
Modeling
Logs
event
history
aggregation
dashboards
Product
Engineering
UX
Stakeholder Customers
DW ETL
Middleware
servletsmodels
Circa 2001: post- big ecommerce successes
“data products”
67Saturday, 31 August 13
68. Workflow
RDBMS
near timebatch
services
transactions,
content
social
interactions
Web Apps,
Mobile, etc.History
Data Products Customers
RDBMS
Log
Events
In-Memory
Data Grid
Hadoop,
etc.
Cluster Scheduler
Prod
Eng
DW
Use Cases Across Topologies
s/w
dev
data
science
discovery
+
modeling
Planner
Ops
dashboard
metrics
business
process
optimized
capacitytaps
Data
Scientist
App Dev
Ops
Domain
Expert
introduced
capability
existing
SDLC
Circa 2013: clusters everywhere
68Saturday, 31 August 13
69. Workflow
RDBMS
near timebatch
services
transactions,
content
social
interactions
Web Apps,
Mobile, etc.History
Data Products Customers
RDBMS
Log
Events
In-Memory
Data Grid
Hadoop,
etc.
Cluster Scheduler
Prod
Eng
DW
Use Cases Across Topologies
s/w
dev
data
science
discovery
+
modeling
Planner
Ops
dashboard
metrics
business
process
optimized
capacitytaps
Data
Scientist
App Dev
Ops
Domain
Expert
introduced
capability
existing
SDLC
Circa 2013: clusters everywhere
“optimize topologies”
69Saturday, 31 August 13
70. Amazon
“Early Amazon: Splitting the website” – Greg Linden
glinden.blogspot.com/2006/02/early-amazon-splitting-website.html
eBay
“The eBay Architecture” – Randy Shoup, Dan Pritchett
addsimplicity.com/adding_simplicity_an_engi/2006/11/you_scaled_your.html
addsimplicity.com.nyud.net:8080/downloads/eBaySDForum2006-11-29.pdf
Inktomi (YHOO Search)
“Inktomi’s Wild Ride” – Erik Brewer (0:05:31 ff)
youtu.be/E91oEn1bnXM
Google
“Underneath the Covers at Google” – Jeff Dean (0:06:54 ff)
youtu.be/qsan-GQaeyk
perspectives.mvdirona.com/2008/06/11/JeffDeanOnGoogleInfrastructure.aspx
MIT Media Lab
“Social Information Filtering for Music Recommendation” – Pattie Maes
pubs.media.mit.edu/pubs/papers/32paper.ps
ted.com/speakers/pattie_maes.html
Primary Sources
70Saturday, 31 August 13
71. Cluster Computing’s Dirty Little Secret
people like me make a good living by leveraging high ROI
apps based on clusters, and so the execs agree to build
out more data centers…
clusters for Hadoop/HBase, for Storm, for MySQL,
for Memcached, for Cassandra, for Nginx, etc.
this becomes expensive!
a single class of workloads on a given cluster is simpler
to manage; but terrible for utilization… various notions
of “cloud” help
Cloudera, Hortonworks, probably EMC soon: sell a notion
of “Hadoop as OS” All your workloads are belong to us
regardless of how architectures change, death and taxes
will endure: servers fail, and data must move
Google Data Center, Fox News
~2002
71Saturday, 31 August 13
72. Three Laws, or more?
meanwhile, architectures evolve toward much, much larger data…
pistoncloud.com/ ...
Rich Freitas, IBM Research
Q:
what kinds of disruption in topologies
could this imply? because there’s
no such thing as RAM anymore…
72Saturday, 31 August 13
73. Topologies
Hadoop and other topologies arose from a need for fault-
tolerant workloads, leveraging horizontal scale-out based
on commodity hardware
because the data won’t fit on one computer anymore
a variety of Big Data technologies has since emerged,
which can be categorized in terms of topologies and
the CAP Theorem
C A
P
strong
consistency
high
availability
partition
tolerance
eventual
consistency
“You can have at most two of these properties for
any shared-data system… the choice of which
feature to discard determines the nature of your
system.” – Eric Brewer, 2000 (Inktomi/YHOO)
cs.berkeley.edu/~brewer/cs262b-2004/PODC-keynote.pdf
julianbrowne.com/article/viewer/brewers-cap-theorem
73Saturday, 31 August 13
74. Some Topologies Other Than Hadoop…
Spark (iterative/interactive)
Titan (graph database)
Redis (data structure server)
Zookeeper (distributed metadata)
HBase (columnar data objects)
Riak (durable key-value store)
Storm (real-time streams)
ElasticSearch (search index)
MongoDB (document store)
ParAccel (MPP)
SciDB (array database)
74Saturday, 31 August 13
75. “Return of the Borg”
consider that Google is generations ahead of
Hadoop, etc., with much improved ROI on its
data centers…
Borg serves as a kind of “secret sauce” for
data center OS, with Omega as its next
evolution:
2011 GAFS Omega
John Wilkes, et al.
youtu.be/0ZFMlO98Jkc
Omega: flexible, scalable schedulers for large compute clusters
Malte Schwarzkopf,Andy Konwinski, Michael Abd-El-Malek, John Wilkes
eurosys2013.tudos.org/wp-content/uploads/2013/paper/Schwarzkopf.pdf
75Saturday, 31 August 13
76. “Return of the Borg”
Omega: flexible, scalable schedulers for large compute clusters
Malte Schwarzkopf,Andy Konwinski, Michael Abd-El-Malek, John Wilkes
eurosys2013.tudos.org/wp-content/uploads/2013/paper/Schwarzkopf.pdf
76Saturday, 31 August 13
77. “Return of the Borg”
Return of the Borg: HowTwitter Rebuilt Google’s SecretWeapon
Cade Metz
wired.com/wiredenterprise/2013/03/google-
borg-twitter-mesos
The Datacenter as a Computer: An Introduction
to the Design ofWarehouse-Scale Machines
Luiz André Barroso, Urs Hölzle
research.google.com/pubs/pub35290.html
77Saturday, 31 August 13
78. Mesos – definitions
a common substrate for cluster computing
heterogenous assets in your data center or cloud
made available as a homogenous set of resources
• top-level Apache project
• scalability to 10,000s of nodes
• obviates the need for virtual machines
• isolation between tasks with Linux Containers (pluggable)
• fault-tolerant replicated master using ZooKeeper
• multi-resource scheduling (memory and CPU aware)
• APIs in C++, Java, Python
• web UI for inspecting cluster state
• available for Linux, Mac OSX, OpenSolaris
78Saturday, 31 August 13
79. Mesos – simplifies app development
CHRONOS SPARK HADOOP DPARK MPI
JVM (JAVA, SCALA, CLOJURE, JRUBY)
MESOS
PYTHON C++
79Saturday, 31 August 13
80. Mesos – data center OS stack
HADOOP STORM CHRONOS RAILS JBOSS
TELEMETRY
Kernel
OS
Apps
MESOS
CAPACITY PLANNING GUISECURITYSMARTER SCHEDULING
80Saturday, 31 August 13
81. Prior Practice: Dedicated Servers
DATACENTER
• low utilization rates
• longer time to ramp up new services
81Saturday, 31 August 13
82. Prior Practice: Virtualization
DATACENTER PROVISIONED VMS
• even more machines to manage
• substantial performance decrease due to virtualization
• VM licensing costs
82Saturday, 31 August 13
83. Prior Practice: Static Partitioning
DATACENTER STATIC PARTITIONING
• even more machines to manage
• substantial performance decrease due to virtualization
• VM licensing costs
• static partitioning limits elasticity
83Saturday, 31 August 13
84. MESOS
Mesos: One Large Pool Of Resources
DATACENTER
“We wanted people to be able to program
for the data center just like they program
for their laptop."
Ben Hindman
84Saturday, 31 August 13
85. What are the costs of Virtualization?
benchmark
type
OpenVZ
improvement
mixed workloads 210%-300%
LAMP (related) 38%-200%
I/O throughput 200%-500%
response time order magnitude
more pronounced
at higher loads
85Saturday, 31 August 13
86. What are the costs of Single Tenancy?
0%
25%
50%
75%
100%
RAILS CPU
LOAD
MEMCACHED
CPU LOAD
0%
25%
50%
75%
100%
HADOOP CPU
LOAD
0%
25%
50%
75%
100%
t t
0%
25%
50%
75%
100%
Rails
Memcached
Hadoop
COMBINED CPU LOAD (RAILS,
MEMCACHED, HADOOP)
86Saturday, 31 August 13
87. Compelling arguments for Data Center OS
• obviates the need forVMs (licensing, adiosVMware)
• provides OS-level building blocks for developing new
distributed frameworks (learning curve, adios Hadoop)
• removes significantVM overhead (performance)
• requires less h/w to buy (CapEx), power and fix (OpEx)
• implies lessVMs, thus less Ops overhead (staff)
• removes the complexity of Chef/Puppet (staff)
• allows higher utilization rates (ROI)
• reduces latency for data updates (OLTP + OLAP on same server)
• reshapes cluster resources dynamically (100’s ms vs. minutes)
• runs dev/test clusters on same h/w as production (flexibility)
• evaluates multiple versions without more h/w (vendor lock-in)
87Saturday, 31 August 13
88. Opposite Ends of the Spectrum, One Substrate
Built-in /
bare metal
Hypervisors
Solaris Zones
Linux CGroups
88Saturday, 31 August 13
89. Opposite Ends of the Spectrum, One Substrate
Request /
Response
Batch
89Saturday, 31 August 13
90. Case Study: Twitter (bare metal / on premise)
“Mesos is the cornerstone of our elastic compute infrastructure –
it’s how we build all our new services and is critical forTwitter’s
continued success at scale. It's one of the primary keys to our
data center efficiency."
Chris Fry, SVP Engineering
blog.twitter.com/2013/mesos-graduates-from-apache-incubation
• key services run in production: analytics, typeahead, ads
• Twitter engineers rely on Mesos to build all new services
• instead of thinking about static machines, engineers think
about resources like CPU, memory and disk
• allows services to scale and leverage a shared pool of
servers across data centers efficiently
• reduces the time between prototyping and launching
90Saturday, 31 August 13
91. Case Study: Airbnb (fungible cloud infrastructure)
“We think we might be pushing data science in the field of travel
more so than anyone has ever done before… a smaller number
of engineers can have higher impact through automation on
Mesos."
Mike Curtis,VP Engineering
gigaom.com/2013/07/29/airbnb-is-engineering-itself-into-a-data-driven...
• improves resource management and efficiency
• helps advance engineering strategy of building small teams
that can move fast
• key to letting engineers make the most of AWS-based
infrastructure beyond just Hadoop
• allowed company to migrate off Elastic MapReduce
• enables use of Hadoop along with Chronos, Spark, Storm, etc.
91Saturday, 31 August 13