[Announcements | Schedule and Readings | Assignments and Quizzes | Syllabus/Policies ]

Reading Questions, Lecture 19 (11/17)

Papers: Eddies, Eddies Overhead

Query processing in domains like the Internet is much more difficult than query processing in traditional databases because the performance of the different operators is fundamentally unpredictable, due to network delays and the difficulty of doing size and cost estimation on remotely located objects. Adaptive query processing, of which Eddies are an example, deals with this unpredictability by allowing the query processor to change query plans during query execution. Eddies are an extreme version of this -- each tuple in an Eddy can be executed using a different plan. The first paper describes the basics of how this works, and the second shows that this can be made to be relatively efficient.

As you read the papers, consider the following questions:

  1. What are the overheads associated with the Eddy method? What are the benefits of the method? When would you expect that an Eddy-like approach to query execution would be the appropriate choice?
  2. How does the routing policy in an Eddy affect the performance of the system? Does the ticket-based approach described in the first paper seem like a good one?
  3. What techniques does the second paper recommend for improving the performance of Eddies?

Samuel Madden (madden at csail dot mit dot edu)
Last modified: Fri Nov 12 21:10:51 EST 2004