Distributed systems, scalability, and consensus
View All Tags
I have always been a big fan of the CAP theorem and its corollary PACELC. There are so many articles that talk about these theorems. However, they often end up making misdirected claims by categorizing systems one way or another.
After much googling, I was not able to find a single article that clearly explained how these rules should be applied in real life. So, here is a way to make it easy for a layperson:
This blog post proposes modifications to the MySQL semi-sync replication process in order to improve the overall consistency and resilience of the system.
This is based on my previous blog post on Flexible Paxos and the related paper Flexible Paxos: Quorum intersection revisited by Howard, Malkhi and Spiegelman.
With systems getting more and more distributed, the Paxos algorithm has been gaining popularity. However, a major drawback with today's configurations is that you cannot run too many Paxos servers in a quorum. The sweet spot seems to be five servers. Those who run only three have to deal with the risk of downtime. Those who run seven or more have to face degraded write performance.
In short, it seems that you have to trade-off between uptime and write performance. However, it is possible to have both if we redefined the quorum rules, and this blog intends to show how.