,
3 tweets,
2 min read
Read on Twitter
@colmmacc Great illustration by Colm. And these failures aren’t just theoretical - we do see them all the time. That is why many correct-looking consensus algorithms actually fail. And this doesn’t even account that most failures in today’s networks are gray 1/
@colmmacc Connectivity doesn’t just get severed, but rather some TCP connections work, while others fail. This is because most data center networks are multipath, that is there are multiple distinct paths between any two nodes 2/
@colmmacc Distinct TCP connections are sent (flowhashed) down distinct paths. And when one path fails, the others are often unimpaired. So you can easily end in a state where your health checks pass, but your nodes cannot exchange data. A very tough failure mode. 3/
