CAP 定理——分区是一个动词

CAP 定理——分区是一个动词
CAP theorem – Partition is a verb

原始链接: http://mmudama.github.io/posts/2026/cap-partitions/

CAP 定理是分布式系统中的核心概念，通常被记作“在一致性、可用性和分区容错性中选择两个”。然而，理解*分区容错性*可能比较困难。许多人最初认为“分区”指的是数据分片，想象系统在特定数据集上发生故障。这种理解是错误的。“分区”在 CAP 定理中是一个*动词*，而不是名词。它描述了系统在节点之间网络连接中断时——当网络“分区”时——继续运行的能力。本质上，分区容错性不是关于*数据在哪里*，而是关于*系统在网络故障期间如何表现*。当发生分割时，系统必须在维护一致性（可能牺牲可用性）或保持可用性（可能牺牲一致性）之间做出选择。它关乎优雅的故障处理，确保即使在通信中断的情况下也能运行，即使某些用户会遇到问题。

这个Hacker News讨论的核心是用户对首页上大量低质量、很可能由AI生成的文章感到沮丧。用户质疑这些“毫无思想”的文章，其特点是重复的句子结构，是如何获得关注的。许多评论者表示难以置信作者真正理解所讨论的概念，认为文章只是简单地从LLM中复制粘贴，而没有理解。讨论集中在文章的具体误解上——是语法问题，还是对分布式系统“分区”这一概念的根本性错误。用户讨论了解决这个问题的方法，主要建议举报这些内容。总体基调是对文章质量的批评，并对平台上未经原创、由AI生成的材料的泛滥表示担忧。

原文

People who’ve spent some time around databases usually have an intuitive sense of ACID, even if the acronym slips their mind. And if you work with distributed systems, you usually have a rough idea of the CAP Theorem, even if the details are fuzzy. The first thing you remember about CAP is probably, “Pick two out of three.” And maybe you can even remember the letters without a search engine &mldr;

Since distributed systems require network hops, it feels right for P to mean “Performance” - but that’s not what it means. The P in CAP stands for Partition Tolerance.

For the longest time, “Partition Tolerance” just wouldn’t stick in my brain. Then something finally clicked. In the CAP theorem, partition isn’t a noun - it’s a verb.

I had thought of the word partition in “Partition Tolerance” as referencing a subset of data. After all, when you partition data stores, you shard them by some attribute - an ID or a geographic location, for example. If you think of “partition” as a noun, you might imagine a system that keeps working for everyone except the people in Maryland, whose data is now offline. Of course that’s hard to remember - that doesn’t sound like a good thing. Certainly not for the people in Maryland.

Turns out, “Partition Tolerance” isn’t about a data partition at all; it’s about the behavior of the network. In this context, the word partition is a verb. It means that when one of the nodes goes AWOL, the system continues operating — though what “operating” looks like depends on the system’s design.

This is a crucial distinction because it shifts the focus from a static state (“partition” as a noun) to a dynamic event (“partition” as a verb). It means the system is designed to handle communication failures gracefully, continuing to operate despite network splits or node isolation.

But what about those poor folks in Maryland? Could they still lose access? Yes, but not because their data is sharded. It’s because when the network broke, the system had to pick a side. Depending on the design, Maryland might even be the only side that stays up. Take that, Virginia!

Partition tolerance isn’t about where your data lives. It’s about what your system does.

CAP 定理——分区是一个动词 CAP theorem – Partition is a verb

CAP 定理——分区是一个动词
CAP theorem – Partition is a verb