No Downtime for Data Conversions: Rethinking Hot Upgrades
Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-09-106. July 2009.
Tudor Dumitraş, Priya Narasimhan
Electrical and Computer Engineering
Carnegie Mellon University
Pittsburgh, PA 15213
Unavailability in enterprise systems is usually the result of planned events, such as upgrades, rather than failures. Major system upgrades entail complex data conversions that are difficult to perform on the fly, in the face of live workloads. Minimizing the downtime imposed by such conversions is a time-intensive and error-prone manual process. We present Imago, a system that aims to simplify the upgrade process, and we show that it can eliminate all the causes of planned downtime recorded during the upgrade history of one of the ten most popular websites. Building on the lessons learned from past research on live upgrades in middleware systems, Imago trades off a need for additional storage resources for the ability to perform end-to-end, enterprise upgrades online, with minimal application-specific knowledge.
FULL TR: pdf