Best Practices Highlights

#1 Partition By Function

• Decouple the Unrelated Functionalities.

• Selling functionality is served by one set of applications, bidding by another, search by yet another.

• 16,000 App Servers in 220 different pools

• 1000 logical databases, 400 physical hosts

#2 Split Horizontally

• Break the workload into manageable units.

• eBay’s interactions are stateless by design

• All App Servers are treated equal and none retains any transactional state

• Data Partitioning based on specific requirements

#3 Avoid Distributed

Transactions

• 2 Phase Commit is a pessimistic approach comes with a big COST

• CAP Theorem (Consistency, Availability, Partition Tolerance). Apply any two at any point in time.

• @ eBay No Distributed Transactions of any kind and NO 2 Phase Commit.

#4 Decouple Functions

Asynchronously

• If Component A calls component B synchronously, then they are tightly coupled. For such systems to

scale A you need to scale B also.

• If Asynchronous A can move forward irrespective of the state of B

• SEDA (Staged Event Driven Architecture)

#5 Move Processing to

Asynchronous Flow

• Move as much processing towards Asynchronous side

• Anything that can wait should wait

#6 Virtualize at All Levels • Virtualize everything. eBay created their on O/R layer for abstraction

#7 Cache Appropriately • Cache Slow changing, read-mostly data, meta data, configuration and static data.