Tag
Cloudflare details how a targeted build planner and new scheduler for its DNS Zone Builder cut peak memory use by 85% and halved the delay between a DNS change and its data reaching Quicksilver, as zone sizes grow toward hundreds of millions of records.
The author argues the agent harness is fundamentally a distributed systems problem and cites DeepSeek's Super-agent writeup, which shows a two-level Router/Worker design beats a single chat session on latency and delegation reliability, while raising long-term memory as the key open problem.
Jepsen 的这份分析报告测试了面向 AI 工作负载的多模型数据库 ParticleDB（1.1.4 至 1.1.18.11.rc-22），发现其宣称的"完全 ACID"并不成立，共报告 96 个问题，包括崩溃、死锁、数据丢失及各类一致性异常，最新的 1.1.19-rc.13 声称已修复全部问题。
The paper proposes ATP (Agent Transfer Protocol), a server-mediated, cross-organization communication protocol for the Agentic Web that enables stable addressing, offline message delivery, and post-hoc source verification for LLM agents via domain-level mediation and signed message envelopes.
GreptimeDB explains its architecture of separating compute and storage via shared object storage in distributed deployments, and describes how it reuses SST data files during repartitioning to avoid migrating historical data wholesale, while also comparing the differences with ClickHouse.
The paper proposes AuditFL, a blockchain-backed clearing layer for open federated learning that keeps training and aggregation off-chain while recording content identifiers and economic state on-chain to make coordinator decisions attributable and auditable, preserving model accuracy even under client dropout and invalid updates on MNIST experiments.
开发者计划于十月使用 Go 构建一个分布式 AI 推理网关，探索 AI 基础设施、负载均衡、智能请求路由、缓存及多推理节点间的容错机制。
GitHub 详述其将 Git 仓库迁移至 Azure 的桥接架构：通过改进 Spokes quorum 的写入序列化方式，把最繁忙的仓库也迁入 Azure，并为无持久状态的副本铺平道路，以支撑 agent 规模的开发流量。
作者开源了一个容器化的多智能体研究与报告生成管线，基于 LangGraph 四阶段状态机（Planner/Researcher/Critic/Writer）、MCP 工具服务器，以及 FastAPI + Celery + Redis + PostgreSQL 的分布式异步架构，并分享了生产化设计经验。
llm-d 是一个开源、Kubernetes 原生的分布式 LLM 推理栈，支持 vLLM、SGLang 等引擎，可在 GPU、TPU、XPU 等异构硬件上提供生产级服务，并宣称大幅提升吞吐量与降低 TTFT。
DeepSeek V4.1 Flash 的大规模推理部署文章，介绍 DeepInfer 结合分布式推理、缓存、负载均衡与投机解码提升效率，DeepSeek 单集群日处理 155 万亿 token，每 GPU 小时收入 38 美元，并开源了大部分推理框架。
Jepsen's analysis of Orkes Conductor 5.3.2 found frequent concurrent task execution, workflow abandonment, and premature completion bugs, prompting Orkes to release version 5.5.6 with patches for all serious safety issues.
一篇关于软件设计哲学的工程博客，论述优雅的系统设计往往来自于主动放弃某些需求（如一致性、并发写、低延迟等），并探讨了在业务与技术之间权衡取舍为何困难。
Gergely Orosz highlights engineer Cindy Sridharan's claim that Claude, when given proper instructions and context via a tailored claude.md, now writes better code than she does almost always.
A user shares reflections on MIT's distributed systems course (6.5840/6.824) versus DDIA, linking to the Spring 2026 schedule covering MapReduce, GFS, Paxos, and Raft, plus a guest lecture by Go creator Russ Cox.
Lingua Franca 是一个开源的协调语言（coordination language），用于构建确定性、多线程、时间敏感和分布式系统，支持 C、C++、Python、TypeScript 和 Rust。它源于 UC Berkeley 等机构数十年的研究，通过 reactor 模型消除竞态条件并内建时序语义。
该论文研究长上下文 LLM 服务中远程 KV 缓存层级的性能瓶颈，提出按并行 rank 分片 MLA KV 存储、异步加载等优化，将 p99 延迟从 83.32 秒降至 0.67 秒，并显著提升预填充吞吐。
NVIDIA 担任 PyTorch Conference 北美站赞助商，公布 12 场演讲与 25 个海报环节，涵盖编译器与 CUDA Graphs、FP4/FP8 低精度训练、分布式系统以及物理与智能体 AI 等方向。会议将于 10 月 20-21 日在圣何塞会议中心举行。
A deep-dive into METR's report on the OpenAI/Hugging Face agent incident, examining how ~1,200 agents self-invented coordination protocols like locks, vetoes, leases, and signatures through a shared package cache, and why classic locking still fails for LLM agents.
一位开发者分享了自己制定的 Agent Infra 六个月学习路径，涵盖 Agent Runtime、分布式系统与可靠性、编排、沙箱、可观测性及生产级 Agent 平台六大板块，并推荐了相关课程与学习资源。