Data Management in the Cloud

1
 Outline
• Motivation
– what is cloud computing?
– what is cloud data management?
• Challenges, opportunities and limitations
– what makes data management in the cloud difficult?
• New solutions
– key/value, document, column family, graph, array, and object databases
– scalable SQL databases
• Application
– graph data and algorithms
– usage scenarios

2
 What is Cloud Computing?
• Different definitions for “Cloud Computing” exist
– http://tech.slashdot.org/article.pl?sid=08/07/17/2117221
• Common ground of many definitions
– processing power, storage and software are commodities that are
readily available from large infrastructure
– service-based view: “everything as a service (*aaS)”, where only
“Software as a Service (SaaS)” has a precise and agreed-upon definition
– utility computing: pay-as-you-go model

3
 Service-Based View on Computing

Client Software

Software
User Interface Machine Interface
(SaaS) End User

Platform
Components Services
(PaaS) Application
Developer

Infrastructure
Computation Network Storage
(IaaS) System
Administrator

Server Hardware
Source: Wikipedia (http://www.wikipedia.org)

4
 Terminology
• Term cloud computing usually refers to both
– SaaS: applications delivered over the Internet as services
– The Cloud: data center hardware and systems software
• Public clouds
– available in a pay-as-you-go manner to the public
– service being sold is utility computing
– Amazon Web Service, Microsoft Azure, Google AppEngine
• Private clouds
– internal data centers of businesses or organizations
– normally not included under cloud computing

5
Based on: “Above the Clouds: A Berkeley View of Cloud Computing”, RAD Lab, UC Berkeley
 Utility Computing
• Illusion of infinite computing resources
– available on demand
– no need for users to plan ahead for provisioning
• No up-front cost or commitment by users
– companies can start small (demand unknown in advance)
– increase resources only when there is an increase in need (demand
varies with time)
• Pay for use on short-term basis as needed
– processors by the hour and storage by the day
– release them as needed, reward conservation
• “Cost associativity”
– 1000 EC2 machines for 1 hour = 1 EC2 machine for 1000 hours

6
Based on: “Above the Clouds: A Berkeley View of Cloud Computing”, RAD Lab, UC Berkeley
 Cloud Computing Users and Providers

7
Picture credit: “Above the Clouds: A Berkeley View of Cloud Computing”, RAD Lab, UC Berkeley
 Virtualization
• Virtual resources abstract from physical resources
– hardware platform, software, memory, storage, network
– fine-granular, lightweight, flexible and dynamic
• Relevance to cloud computing
– centralize and ease administrative tasks
– improve scalability and work loads
– increase stability and fault-tolerance
– provide standardized, homogenous computing platform through
hardware virtualization, i.e. virtual machines

8
 Spectrum of Virtualization
• Computation virtualization
– Instruction set VM (Amazon EC2, 3Tera)
– Byte-code VM (Microsoft Azure)
– Framework VM (Google AppEngine, Force.com)
• Storage virtualization
• Network virtualization

Lower-level, Higher-level,
Less management More management

EC2 Azure AppEngine Force.com

9
Slide Credit: RAD Lab, UC Berkeley
 10
Table credit: “Above the Clouds: A Berkeley View of Cloud Computing”, RAD Lab, UC Berkeley
 Economics of Cloud Users
• Pay by use instead of provisioning for peak

Static data center Data center in the cloud

Capacity

Resources
Resources

Capacity

Demand Demand
Time Time

Unused resources

11
Slide Credit: RAD Lab, UC Berkeley
 Economics of Cloud Users
• Risk of over-provisioning: underutilization

Static data center

Capacity
Resources

Demand
Time

Unused resources

12
Slide Credit: RAD Lab, UC Berkeley
 Economics of Cloud Users
• Heavy penalty for under-provisioning

Resources
Capacity

Demand
1 2 3
Time (days)

Lost revenue Lost users

Resources
Resources

Capacity Capacity

Demand Demand
1 2 3 1 2 3
Time (days) Time (days)

13
Slide Credit: RAD Lab, UC Berkeley
 Economics of Cloud Providers
Cost in Medium Cost in Very Large
Resource Data Center Data Center Ratio

Network $95/Mbps/month $13/Mbps/month 7.1x

Storage $2.20/GB/month $0.40/GB/month 5.7x
Administration ≈140 servers/admin >1000 servers/admin 7.1x
Source: James Hamilton (http://perspectives.mvdirona.com)
• Cloud computing is 5-7x cheaper than traditional in-house
computing
• Power/cooling costs: approx double cost of storage, CPU, network
• Added benefits (to cloud providers)
– utilize off-peak capacity (Amazon)
– sell .NET tools (Microsoft)
– reuse existing infrastructure (Google)

14
Slide Credit: RAD Lab, UC Berkeley
 What is Cloud Data Management?
• Data management applications are potential candidates for
deployment in the cloud
– industry: enterprise database system have significant up-front cost that
includes both hardware and software costs
– academia: manage, process and share mass-produced data in the cloud
• Many “Cloud Killer Apps” are in fact data-intensive
– Batch Processing as with map/reduce
– Online Transaction Processing (OLTP) as in automated business
applications
– Online Analytical Processing (OLAP) as in data mining or machine
learning

15
 Scientific Data Management Applications
• Old model
– “Query the world”
– data acquisition coupled to a specific hypothesis
• New model
– “Download the world”
– data acquired en masse, in support of many hypotheses
• E-science examples
– astronomy: high-resolution, high-frequency sky surveys, …
– oceanography: high-resolution models, cheap sensors, satellites, …
– biology: lab automation, high-throughput sequencing, ...

16
Slide Credit: Bill Howe, U Washington
 Scaling Databases
• Flavors of database scalability
– lots of (small) transactions
– lots of copies of the data
– lots of processors running on a single query (compute intensive tasks)
– extremely large data set for one query (data intensive tasks)
• Data replication
– move data to where it is needed
– managed replication for availability and reliability

17
 Revisit Cloud Characteristics
• Compute power is elastic, but only if workload is parallelizable
– transactional database management systems do not typically use a
shared-nothing architecture
– shared-nothing is a good match for analytical data management
– some things parallelize well (i.e. sum), some do not (i.e. median)
– Think about: Google gmail, Amazon web site – easy? Difficult?
– Google App Engine – API forces ability to run in shared nothing
• Scalability
– in the past: out-of-core, works even if data does not fit in main memory
– in the present: exploits thousands of (cheap) nodes in parallel

18
Based on: “Data Management in the Cloud: Limitations and Opportunities”, IEEE, 2009.
 Parallel Database Architectures
Shared nothing Shared disc Shared memory

interconnect …

… interconnect

… interconnect

processor memory disk

Source: D. DeWitt and J. Gray: “Parallel Database Systems: The Future of

19
High Performance Database Processing”, CACM 36(6), pp. 85-98, 1992.
 Revisit Cloud Characteristics
• Data is stored at an untrusted host
– there are risks with respect to privacy and security in storing
transactional data on an untrusted host
– particularly sensitive data can be left out of analysis or anonymized
– sharing and enabling access is often precisely the goal of using the
cloud for scientific data sets
– where exactly is your data? and what are that country’s laws?

20
Based on: “Data Management in the Cloud: Limitations and Opportunities”, IEEE, 2009.
 Revisit Cloud Characteristics
• Data is replicated, often across large geographic distances
– it is hard to maintain ACID guarantees in the presence of large-scale
replication
– full ACID guarantees are typically not required in analytical applications
• Virtualizing large data collections is challenging
– data loading takes more time than starting a VM
– storage cost vs. bandwidth cost
– online vs. offline replication

21
Based on: “Data Management in the Cloud: Limitations and Opportunities”, IEEE, 2009.
 Challenges
• Trade-off between functionality and operational cost
– restricted interface, minimalist query language, limited consistency
guarantees
– pushes more programming burden on developers
– enables predictable services and service level agreements
• Manageability
– limited human intervention, high-variance workloads, and a variety of
shared infrastructures
– need for self-managing and adaptive database techniques

22
Based on: “The Claremont Report on Database Research”, 2008
 Challenges
• Scalability
– today’s SQL databases cannot scale to the thousands of nodes deployed
in the cloud context
– hard to support multiple, distributed updaters to the same data set
– hard to replicate huge data sets for availability, due to capacity (storage,
network bandwidth, …)
– storage: different transactional implementation techniques, different
storage semantics, or both
– query processing and optimization: limitations on either the plan
space or the search will be required
– programmability: express programs in the cloud

23
Based on: “The Claremont Report on Database Research”, 2008
 Challenges
• Data privacy and security
– protect from other users and cloud providers
– specifically target usage scenarios in the cloud with practical incentives
for providers and customers
• New applications: “mash up” interesting data sets
– expect services pre-loaded with large data sets, stock prices, web
crawls, scientific data
– data sets from private or public domain
– might give rise to federated cloud architectures

24
Based on: “The Claremont Report on Database Research”, 2008
 Transactional Data Management – Cloud or not?
• Transactional Data Management
– Banking, airline reservation, e-commerce, etc…
– Require ACID, write-intensive
• Features
– Do not typically use shared-nothing architectures (changing a bit)
– Hard to maintain ACID guarantees in the face of data replication over
large geographic distances
– There are risks in storing transactional data on an untrusted host
• Conclusion: not appropriate for the cloud

25
Based on: “Data Management in the Cloud: Limitations and Opportunities”, IEEE, 2009.
 Analytical Data Management – Cloud or not?
• Analytical Data Management
– Query data from a data store for planning, problem solving, decision
support
– Large scale
– Read-mostly
• Features
– Shared-nothing architecture is a good match for analytical data
management (Teradata, Greenplum, Vertica…)
– ACID guarantees typically not needed
– Particularly sensitive data left out of analysis
• Conclusion: appropriate for the cloud

26
Based on: “Data Management in the Cloud: Limitations and Opportunities”, IEEE, 2009.
 Cloud DBMS Wish List
• Efficiency
• Fault tolerance (query restart not required, commodity hw)
• Heterogeneous environment (performance of compute nodes
not consistent)
• Operate on encrypted data
• Interface with business intelligence products

27
Based on: “Data Management in the Cloud: Limitations and Opportunities”, IEEE, 2009.
 Option 1: MapReduce-like software
• Fault tolerance (yes, commodity hw)
• Heterogeneous environment (yes, by design)
• Operate on encrypted data (no)
• Interface with business intelligence products (no, not SQL-
compliant, no standard)
• Efficiency (up for debate)
– Questionable results in the MapReduce paper
– Absence of a loading phase (no indices, materialized views)

28
Based on: “Data Management in the Cloud: Limitations and Opportunities”, IEEE, 2009.
 Option 2: Shared-Nothing Parallel Database
• Interface with business intelligence products (yes, by design)
• Efficiency (yes)
• Fault tolerance (no - query restart required)
• Heterogeneous environment (no)
• Operate on encrypted data (no)

29
Based on: “Data Management in the Cloud: Limitations and Opportunities”, IEEE, 2009.
 Option 3: A Hybrid Solution
• HadoopDB (http://db.cs.yale.edu/hadoopdb/hadoopdb.html)
– A hybrid of DBMS and MapReduce technologies that targets analytical
workloads
– Designed to run on a shared-nothing cluster of commodity machines,
or in the cloud
– An attempt to fill the gap in the market for a free and open source
parallel DBMS
– Much more scalable than currently available parallel database systems
and DBMS/MapReduce hybrid systems.
– As scalable as Hadoop, while achieving superior performance on
structured data analysis workload
• Commercialized as Hadapt (hadapt.com)

30
 Why NoSQL?
• Value of relational databases
– Persistent data
– Concurrency/transactions
– Integration
– (Mostly) Standard Model
• Impedance Mismatch
• Application vs. Integration databases

31
Based on: “NoSQL Distilled: A Brief Guide to the Emerging World of Polyglot Persistence”, 2013
 Attack of the Clusters
• Data growth (links, social networks, logs, users)
• Need to scale to accommodate growth
• Traditional RDBMS (Oracle / Microsoft SQL Server) – shared
disk – don’t scale well
• “technical issues are exacerbated by licensing costs”
– Google, Amazon influential
• “The interesting thing about Cloud Computing is that we’ve
redefined Cloud Computing to include everything that we
already do… I don’t’ understand what we would do differently
in the light of Cloud Computing other than change the wording
of some of our ads.” – Larry Ellison

Based on: “NoSQL Distilled: A Brief Guide to the Emerging World of Polyglot Persistence”, 2013
32
 Emergence of NoSQL
• No strong definition, but…
– Do not use SQL
– Typically open-source
– Typically oriented towards clusters (but not all)
– Operate without a schema
• Various types (in order of complexity)
– Key-value stores
– Document Stores
– Extensible Record Stores

Based on: “NoSQL Distilled: A Brief Guide to the Emerging World of Polyglot Persistence”, 2013
33
 What were we talking about?
• Cloud Computing
– Utility Computing
– Virtualization
– Economics (pay as you go)
• Data management in the cloud
– Cloud characteristics (elasticity if parallelizable, untrusted host, large
distances)
– Transactional vs. Analytical
– Wish List
– Map Reduce vs. Shared-Nothing -> Hybrid
• DB vs. NoSQL in two lines…
– Database: complex / concurrent
– NoSQL: simple / scalable

34
 References
• M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. H. Katz, A. Konwinski,
G. Lee, D. A. Patterson, A. Rabkin, I. Stoica, M. Zaharia: Above the
Clouds: A Berkeley View of Cloud Computing. Tech. Rep. No.
UCB/EECS-2009-28, 2009.
• D. J. Abadi: Data Management in the Cloud: Limitations and
Opportunities. IEEE Data Eng. Bull. 32(1), pp. 3—12, 2009.
• R. Agrawal, A. Ailamaki, P. A. Bernstein, E. A. Brewer, M. J. Carey, S.
Chaudhuri, A. Doan, D. Florescu, M. J. Franklin, H. Garcia‐ Molina, J.
Gehrke, L. Gruenwald, L. M. Haas, A. Y. Halevy, J. M. Hellerstein, Y. E.
Ioannidis, H. F. Korth, D. Kossmann, S. Madden, R. Magoulas, B. Chin
Ooi, T. O’Reilly, R. Ramakrishnan, S. Sarawagi, M. Stonebraker, A. S.
Szalay, G. Weikum: The Claremont Report on Database Research.
2008.
• P. Sadalage, M. Fowler. NoSQL Distilled: A Brief Guide to the
Emerging World of Polyglot Persistence. 2013
35