# Quorum Availability DevFeed: [Quorum Availability]() Original publisher: [Read original article]() Author: Marc Brooker Published: 2021-01-06T00:00:00Z Content type: tutorial Language: en Sources: [Marc Brooker's Blog](), [Marc Brooker's Blog]() Topics: [Databases](), [systems](), [math](), [datacenter](), [Availability]() Tags: [availability](), [databases](), [datacenter](), [math](), [systems]() ## AI overview This article explains the counterintuitive availability behavior of quorum systems with different cluster sizes. Using Paxos clusters, majority quorums, and independently failing machines, it shows that larger clusters are more available when fewer than half the machines fail, while smaller clusters can be more available when failure rates exceed 50%. ## Source excerpt Quorum Availability It's counterintuitive, but is it right? In our paper Millions of Tiny Databases, we say this about the availability of quorum systems of various sizes: As illustrated in Figure 4, smaller cells offer lower availability in the face of small numbers of uncorrelated node failures, but better availability when the proportion of node failure exceeds 50%. While such high failure rates are rare, they do happen in practice, and a key design concern for Physalia. And this is what Figure 4 looks like: The context here is that a cell is a Paxos cluster, and the system needs a majority quorum for the cluster to be able to process requests1. A cluster of one box needs one box available, five boxes need three available and so on. The surprising thing here is the claim that having smaller clusters is actually better if the probability of any given machine failing is very high. The paper doesn't explain it well, and I've gotten a few questions about it. This post attempts to do better. Let's start by thinking about what happens for a cluster of one machine (n=1), in a datacenter of N machines (for very large N). We then fail each machine independently with probability p. What is the probability that our one machine failed? That's trivial: it's p. Now, let's take all N machines and put them into a cluster of n=N. What's the probability that a majority of the cluster is available? For large N, it's 1 for p < 0.5, and 0 for p > 0.5. If less than half the machines fail, less than half have failed. If more than half the machines fail, more than half have failed. Ok? Notice how a cluster size of 1 is worse than N up until p = 0.5 then better after. Peleg and Wool say: ... for 0 < p < ½ the most available NDC2 is shown to be the "democracy" (namely, the minimal majority system), while the "monarchy" (singleton system) is least available. Due to symmetry, the picture reverses for ½ < p < 1. Here, the minimal majority system is the one I'd call a majority quorum, and is us