Tag
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。
一篇工程博客分享了如何通过将 Unreal 编辑器着色器哈希从缓慢的 SHA1 替换为引擎内已有的 xxHash，从而加速编辑器启动与资产处理。文章强调应避开已被高度优化的代码路径，转而寻找被忽视的低垂果实。
Ai2 分享其 GPU 调度器的工程实践，通过 GPU 时间预算、分层公平共享分配和时间片机制取代优先级调度器，将其最大 H100 集群的中位排队等待时间从 5 分钟降至 24 秒。
Gergely Orosz highlights turbopuffer publicly documenting its in-progress rewrite of the tpuf storage engine (v3), grinding query plans in public rather than only celebrating after the rewrite is finished and fast.
Datadog rebuilt their Git serving infrastructure with a custom mirror called gitretriever, handling 20 times more CI traffic while maintaining low latency and reducing CPU usage.
Manifest explains why it deprecated its LLM router, arguing that model routing introduces unpredictability, breaks behavior consistency, and that prompt complexity cannot be inferred from the prompt alone, making caching and deliberate model selection more effective for most use cases.
Anthropic published an engineering blog post detailing a multi-agent system, using Claude Opus 4 as the main orchestrator and Claude Sonnet 4 as sub-agents. The multi-agent system improved performance by 90.2% over a single Claude Opus 4, while token consumption increased by approximately 15x. It also summarized five collaboration patterns.
The author details the process of designing a custom query language tailored for non-technical analysts to filter vehicle maintenance data, outlining user needs, data schema, and specific use cases.