Tag
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。
The celld team describes how they use deterministic simulation testing (DST) to explore event orderings and reproduce hard-to-reproduce bugs in their Cloudflare Workers/Durable Objects runtime, including how a seeded simulator controls scheduling, storage faults, and time to verify system invariants.
Jane Street 分享了其内部消息框架 Aria 的扩展实践：通过为 topic partition 引入索引与树拆分策略优化 tip recovery 过滤流程，在生产负载上实现 30% 的 CPU 使用率下降。
WorldCast 提出了一种分布式多人世界模型，每个玩家本地运行视频生成器与状态模型，通过交换玩家状态和共享场景状态实现一致、可扩展的实时多人生成，在 Counter-Strike 2 上验证了其一致性与可扩展性。
作者指出多智能体流水线最常见的失败模式并非单个智能体出错，而是两个各自正确的智能体对'完成'的定义不一致导致交接契约失效；通过显式定义每个交接的保证条件（而非仅数据格式）修复了大量输出质量问题。
A comprehensive, bookmark-worthy roadmap covering software architecture from fundamentals and design principles to distributed systems, DDD, evolution, and communication practices for developers aiming to become architects.
A FAST 2018 research paper proposing protocol-aware recovery techniques for consensus-based storage systems that efficiently recover state after failures without violating correctness.
Gergely Orosz reveals that Uber now operates fewer microservices in 2026 than it did in 2020, continuing the industry shift toward well-sized "macroservices" to reduce maintenance overhead. More details will come in an upcoming podcast with Uber engineer Uday Kiran Medisetty.
该论文形式化了分布式多智能体委派中的容错预算守恒机制，提出基于量化资源向量、排他托管信用与委托DAG的防超额花费方案，并通过TLA+模型检查与崩溃注入实验证明了账本与效果守恒等安全属性。
Cloudflare announced K2 in public beta, a durable, fully serverless event streaming primitive on its Developer Platform that implements a partitioned durable log on top of R2 object storage, enabling scalable event decoupling similar to Kafka but designed for the edge.
Tigris details migrating async tasks from a FoundationDB-based QuiCK-style queue to Kafka to reduce read/write load and custom code, while keeping some queues in FoundationDB to avoid the dual-write problem.
Harvard's open-source textbook *Machine Learning Systems*, derived from CS249R, covers building and deploying ML systems — from single-machine setups to distributed training, data, deployment, and monitoring — and ships with TinyTorch, a ~20-module framework for building a minimal trainable system from tensors up.
The article introduces a distributed benchmarking website for AI agents, allowing community members to test and share results across various local models and hardware configurations.
A tweet describing Modal as a planet-scale computer built from warehouse-scale computers, highlighting its distributed computing capabilities.
A call for collaborators to research the verification of AI agent actions, focusing on ensuring intended state changes through authorized paths without unintended side effects.
The author proposes a local-first, distributed model for AI agents where users own persistent agents that can operate locally, negotiate with other agents, and interact across environments with structured actions and an agent economy.
This paper synthesizes research on efficient distributed LLM serving, connecting vLLM and llm-d as complementary layers and proposing an Inference Execution Planner for future scheduler development.
Delta is Meta's object storage service that uses chain replication to achieve high availability and strong consistency, designed for critical bootstrap and disaster recovery workloads.
The article explores a search for a compositional theory of self-stabilization in distributed systems, critiquing recent papers on metastable failures and formal methods like rely-guarantee contracts.
An analysis of agent task queues reveals that duplicate claims by different agents can both be verified but only one is counted, leading to invisible waste and issues with self-reported timestamps and lack of exclusion events.