The document discusses cloud computing and designing applications for scalability and availability in the cloud. It covers key considerations for moving to the cloud like design for failure, building loosely coupled systems, implementing elasticity, and leveraging different storage options. It also discusses challenges like application scalability and availability and how to address them through patterns like caching, partitioning, and implementing elasticity. The document uses examples like MapReduce to illustrate how to build applications that can scale horizontally across infrastructure in the cloud.

1. Cloud Computing

2. Agenda
 Design Considerations in moving to Cloud
 Introduction to Windows Azure
 Demo

4. Business Benefits of Cloud Computing
 Almost zero upfront infrastructure investment.
 Just-in-time infrastructure.
 More efficient resource utilization.
 Usage based costing.
 Reduced time to Market.

5. Cloud Computing

6. Cloud Computing

7. Challenges Faced by Apps in the
Cloud
 Application Scalability
 Cloud promises rapid (de)provisioning of resources.
 How do you tap into that to create scalable systems?
 Application Availability
 Underlying resource failures happen
… usually more frequently than in
traditional data centers.
 How do you overcome that to create highly available systems?

8. The Scalability Challenge
 Two different components to scale:
 State (inputs, data store, output)
 Behavior (business logic)
 Any non-trivial application has both.
 Scaling one component means scaling the other, too.

9. Scalability Considerations
 Performance vs Scalability
 Latency vs Throughput
 Availability vs Consistency
How do you manage overload ?

10. Scalable Service
 A scalable architecture is critical to take advantage of scalable
infrastructure.
 Characteristic of Scalable Service:
 Increasing resources results in a proportional increase in performance
 A scalable service is capable of handling heterogeneity
 A scalable service is operationally efficient
 A scalable service is resilient
 A scalable service becomes more cost effective when it grows.

11. 1. Design for Failure
and nothing will really fail
 Avoid single points of failure
 Assume everything fails and design backwards.
 Applications should continue to function even if the physical hardware fails
or removed or replaced.

12. Design for Failure contd..
 Unit of failure is a single host
 Where possible, choose services and infrastructure that assume host failures
happen.
 By building simple services composed of a single host, rather then multiple
dependent hosts, one can create replicated service instances that can
survive host failures.
 Make your services small and stateless.
 Relax consistency requirements.

13. 2. Build Loosely Coupled Systems
The looser they’re coupled, the bigger they scale
 Loosely Coupled Dependencies
 Avoid complex design and interactions.
 Best Practices:
 Tiered Architecture
 Scale out units
 Single role

14. 3. Implement Elasticity
Elasticity is fundamental property of cloud
 Ability to add and remove capacity as and when it is required.
 Use Elastic Load Balancing.
 Use Auto-Scaling (free)

15. 4. Build Security in every layer
Design with security in mind
 Encrypt data at rest.
 Encrypt data at transit (SSL)
 Consider encrypted file system for sensitive data.
 Rotate your credentials, Pass in arguments encrypted.
 Use MultiFactor authentication.
 Restrict external access to specific IP ranges.

16. 5. Don’t Fear Constraints
Re-think architectural constraints
 More RAM? – Distribute load across machines. Shared distributed cache.
 Better IOPS on database – Multiple read-only / Sharding.
 Performance – Caching at different levels

17. 6. Think Parallel
Serial and Sequential is now history
 Experiment different architectures in parallel
 Multi-threading and Concurrent requests to cloud services.
 Run parallel Map Reduce Jobs
 Use Elastic Load balancing to distribute load across multiple servers.
 Decompose job into its simplest form – and with “Shared Nothing”

18. 7. Leverage many storage options
One Size does not fit all
 Amazon S3 / Azure Blob – Large Static Objects
 Amazon Simple DB / Azure Tables – Data indexing and Querying
 Amazon RDS / SQL Azure – RDMBS Service – Automated and Managed
MySQL/Azure
 Amazon Cloud Front / Azure CDN – Content Distribution

19. Cloud Architecture Lessons
 Design for failure and nothing fails.
 Loose coupling sets you free.
 Implement Elasticity.
 Build Security in every layer.
 Don’t fear constraints.
 Think Parallel.
 Leverage many storage options.

20. Windows

21. Why Windows Azure?

22. Azure Overview

23. Execution Models

24. Application Scenarios

25. Data Management

26. Storage: What are our options?

27. Blob Storage Concepts

28. Networking

29. Business Analytics

30. Messaging

31. Caching

32. Stages of Service Deployment

33. Packaging & Deployment

34. Questions

35. References

36. http://WindowsAzure.com

37. MISC SLIDES

38. App Scalability Patterns for State
 Data Grids  CAP theorem: Data Consistency
 Distributed Caching  Eventually Consistent
 HTTP Caching  Atomic Data
 Reverse Proxy  DB Strategies
 CDN  RDBMS
 Denormalization
 Concurrency
 Sharding
 Message-Passing
 NOSQL
 Dataflow
 Key-Value store
 Software Transactional Memory
 Document store
 Shared-State  Data Structure store
 Partitioning  Graph database

39. App Scalability Patterns for Behavior
 Compute Grids  Load Balancing
 Round-robin
 Event-Driven Architecture
 Random
 Messaging
 Weighted
 Actors
 Dynamic
 Enterprise Service Bus
 Parallel Computing
 Domain Events  Master/Worker
 Event Stream Processing  Fork/Join
 Event Sourcing  MapReduce
 Command & Query Responsibility  SPMD
Segregation (CQRS)  Loop Parallelism

40. The Availability Challenge
 Availability: Tolerate failures
 Traditional IT focuses on increasing MTTF
 Mean Time to Failure
 Cloud IT focuses on reducing MTTR
 Mean Time to Recovery

41. Data modelling
 Classic distributed systems focused on ACID semantics
 Atomicity: either the operation (e.g., write) is performed on all
replicas or is not performed on any of them
 Consistency: after each operation all replicas reach the same state
 Isolation: no operation (e.g., read) can see the data from another
operation (e.g., write) in an intermediate state
 Durability: once a write has been successful, that write will persist
indefinitely
 Modern Internet Systems – focused on BASE
 Basically Available
 Soft-state (or scalable)
 Eventually consistent

42. CAP Theorem
Any distributed system has three properties – CAP
 Strong Consistency: all clients see the same view, even in the presence of
updates
 High Availability: all clients can find some replica of the data, even in the
presence of failures
 Partition-tolerance: the system properties hold even when the system is
partitioned
As per CAP theorem you can only have two of these three properties. Choice
of which feature to discard determines the nature of your system.

43. Map Reduce
 Model for processing large data sets.
 Many tasks composed of processing lots of data to produce lots of other
data
 Want to use hundreds or thousands of CPUs... but this needs to be easy!
 Contains Map and Reduce functions.

44. Programming model
 Input & Output: each a set of key/value pairs
 Programmer specifies two functions:
 map (in_key, in_value) -> list(out_key, intermediate_value)
 Processes input key/value pair
 Produces set of intermediate pairs
 reduce (out_key, list(intermediate_value)) -> list(out_value)
 Combines all intermediate values for a particular key
 Produces a set of merged output values (usually just one)

45. So How Does It Work?

46. So How Does It Work?

48. Example
 Page 1: the weather is good
 Page 2: today is good
 Page 3: good weather is good.

49. Map output
 Worker 1:
 (the 1), (weather 1), (is 1), (good 1).
 Worker 2:
 (today 1), (is 1), (good 1).
 Worker 3:
 (good 1), (weather 1), (is 1), (good 1).

50. Reduce Input
 Worker 1:
 (the 1)
 Worker 2:
 (is 1), (is 1), (is 1)
 Worker 3:
 (weather 1), (weather 1)
 Worker 4:
 (today 1)
 Worker 5:
 (good 1), (good 1), (good 1), (good 1)

51. Reduce Output
 Worker 1:
 (the 1)
 Worker 2:
 (is 3)
 Worker 3:
 (weather 2)
 Worker 4:
 (today 1)
 Worker 5:
 (good 4)

52. Conclusion – Map Reduce
 MapReduce has proven to be a useful abstraction
 Greatly simplifies large-scale computations
 Fun to use: focus on problem, let library deal w/ messy details