Week 2, Day 2
A structure for Software as a Service (SaaS)
Source: Condos and clouds.
- Lowest level: Platform software providing basic services (local storage, processes, synchronization, networking)
-
Separate out common features that are above the operating system but more general than one application
- Cavage calls this infrastructure “plumbing” and “concierge” (two names for the same thing)
- The term we prefer (from the Google Datacenter people) is cluster-level infrastructure
- Neither term is precise—people may have different opinions about what lies at this level
-
The cluster-level infrastructure can deal with new features of cloud technology:
- Rapid scalability: provisioning new servers on demand.
- Failure tolerance: Datacenters use hardware that is relatively unreliable; software also fails.
- Rapid changes in datacenter networking: Processes can move to different machines, new servers can be added to an application, …
- Transmission delays due to geography and network outages (the Consistency, Availability, and Partioning problem—CAP)
-
Application-level software can provide commonly-required tools
- Storage
- Credit card payment processing
- Sending emails
- User ID management and login
Breaking one service into multiple services
-
A top-level user service, such as “Display the page for The 2009-2014 World Outlook for 60-Milligram Containers of Fromage Frais”, will be composed of many service calls within the application:
- Add this book to the “recently-visited” session data for this user
- Get sample pages, if any
- Check session identifier remains valid
- Get number of items in the shopping cart
-
Each service, in turn, calls other services:
Source: Condos and clouds, Fig. 4. Copyright 2013, ACM.
- Each lower level has tighter SLA (response time)
Ensuring scalability
- How can we take advantage of elastic computing?
-
If there’s a sudden surge of requests, how do we scale up quickly?
- Minimize communication between different user requests
- Have multiple machines ready to serve a given user request
- Sharding (Helland calls it “partitioning”) breaks the database in sections with separate keys
- Replication creates multiple copies of each shard
These methods are also listed in the table from Section 2.2 of The Datacenter as a Computer, 2d. Ed. that you read for last Wednesday.
Source: Condos and clouds, Fig. 4. Copyright 2013, ACM.