|
|
Traditional HA, high volume transactional
systems have enjoyed the safe harbor of a private network. This has
provided relative predictability of the workload and the ability to
manage that workload as necessary. The profile of applications can
be engineered and understood. Growth, peak demands, and customer behavior
can be analyzed and prepared for.
The workload faced by today's largest internet commerce engines is not
predictable. Each minute, day, and season at Amazon.com surfaces new,
interesting challenges to the design assumptions. Customers respond to
external events and Amazon promotions in unexpected ways. Customers and
non-customers also participate in innovating the technical environment,
sometimes with the sole objective of breaking the system.
Amazon.com has chosen design patterns that fit this hostile environment. The
developing architecture seeks rapid innovation, unconstrained scalability,
always-on availability, and low marginal cost of ownership. Amazon.com's
technology platform is in an early stage, yet signficant wins have been
achieved. More importantly, Amazon is learning to build long-term scalability
while feature code is rolling to production every
hour.
In today’s world, applications span autonomous computer systems. These systems do not trust each other. This talk examines the consequences of this distrust and introduces the Autonomous Computing model for application design. We argue that this subsumes client server as a model for computing.
First, we introduce the notion of a fiefdom as an independent computing environment that refuses to trust any outsiders and maintains tight control over a set of mission critical data. Next, we describe another type of computing component called an emissary. Emissaries help prepare requests to submit to a fiefdom. They operate exclusively on snapshot reference data and single-user data. We spend a bunch of time on the usage of snapshot reference data and how work can be distributed.
Big Data: In search of Petabyte Databases, Jim Gray & Tony Hey [Chair: Pat Helland]
The panel discussion is on the problems encountered in building scalable, highly available, and cost efficient systems for large retail businesses. The focus will be on where current technologies have not effectively solved some of the problems. Retail businesses fall into three broad categories -- traditional brick-and-mortar, catalogue based, and e-commerce. This panel will include members from all three segments. The outcome of the panel discussion will result in highlighting two or three burning issues that are central to large retail businesses but are not effectively solved by existing infrastructure software.