CS262B Reading Summary

Lessons from Giant-Scale Services

Eric A. Brewer

Summary by Feng Zhou
2/18/2004

Strong points of the paper are:

  1. Differenciate "yield" and "harvest" gives insight into the unique properties of Internet services.  They are both quality-of-service metrics.  But traditionally, shedding requests is the often used method to cope with overloading.  However, here the author points out that because the properties of Internet services, in a lot of times returning results based on partial data may be better than fulfilling part of the requests.  At least nobody sees error messages and he/she gets some useful info.
  2. In general, the DQ principle is useful when designing and provisioning a Internet service.  It helps thinking about the impact on performance and availability of various  system components, in a simple straightforward way.  It also helps planning fail-over strategies when certain components go down.
  3. The paper provides useful facts about giant-scale Internet service.  E.g., the peak-to-average ratio is from 1.6:1 to 6:1; various data about several clusters used in Internet services.
One major flaw.

The DQ principle has one limitations in application.  It only applies to those services who's resource usage is roughly proportional to the amount of data processed.  In some services, the amount of resource spent on each data item may vary greatly according to the request type.