The blog post mentioned correlated failures in an availability zone. You likely could reduce this a bit, but still run into this frequently enough