Consistency is not a localized property
Summary
一篇关于分布式系统一致性的技术博客,指出一致性并非单个组件的局部属性,即使 Apache Kafka 花费数年实现 exactly-once 语义,仍需要应用端到端设计和专人对整体属性负责。
View Cached Full Text
Cached at: 10/11/26, 01:15 AM
Similar Articles
Exactly-once execution doesn't exist, and agent stacks need to accept that
This article argues that achieving exactly-once execution in distributed systems is impossible, and that agent stacks must be designed to deal with this limitation.
Data liberation: Apache Kafka's native cluster mirroring
Apache Kafka introduces native cluster mirroring to embed cross-cluster replication directly into the broker, simplifying data movement between clusters with byte-for-byte transfer and exact offset preservation.
Retries don't fix eventual consistency
The article argues that retrying messages in eventually consistent distributed systems treats normal state as failure, and instead suggests storing incoming data and processing when all prerequisites arrive, eliminating the need for retries and dead-letter queues.
@vintcessun: The barrier to developing multi-agent systems is too high; those who haven't studied Agent theory dare not touch it. As a result, project implementation is difficult, and teams can only rely on a few experts. This paper directly takes mature architectural patterns from distributed systems (publish-subscribe, message queues, etc.) and defines a minimal set of Agent concepts mapped onto them. Even students with no DS experience can use it...
This paper proposes directly mapping mature architectural patterns from distributed systems (such as publish-subscribe and message queues) to multi-agent systems to lower the development barrier. It was validated in a course: even students with no distributed systems experience could get started with gRPC and RabbitMQ, achieving an average score above 80%.
@siddontang: A few days ago, someone asked me: What Multi-Agent System is TiDB using internally now? I answered very directly: Actually, we are deliberately avoiding Multi-Agent. If one Agent can handle it, let one Agent do it. If it really needs to be split, at most put a Coordinator above…
The TiDB team deliberately avoids using Multi-Agent systems internally, based on lessons learned from distributed systems and referencing Anthropic's related research, emphasizing that optimizing a single Agent might be more effective.