Mike Christian interviewed at Velocity 2012
summary
Yahoo builds availability around the assumption that entire metros fail, replicating data so services run from multiple locations and accepting small latency hits during failover. Operation uses DNS-based failure detection and automated traffic shifting, and it takes about 8 minutes to shift an entire data center's traffic. Planning focuses on assume you'll lose an entire metro region rather than perfect within-rack redundancy, with pragmatic tradeoffs like using commercial power in lab sites and shifting traffic when UPS or power issues occur.
AI-written summary · about
