Let's think about that cache again: when that one node is down, overloaded, busy being deployed, etc that object is not available. This property can make operating high hit-rate caches and storage systems particularly difficult: any kind of deployment or change can look to clients like a kind of rolling outage. However, with erasure coding, single failures (or indeed any M−k number of failures) have no availability impact.

Erasure Coding versus Tail Latency
from marcbrooker@gmail.com (Marc Brooker) favicon