← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 7010.l · 生成模型:DeepSeek-V4.1-Flash

在线服务的五类代价

理解等级:高级

在线服务的代价:其一必须有延迟预算——模型要访问全部相关特征,新查询得在毫秒内转成特征并取回别处存的特征,凡单个训练样本所需数据无法随 API 调用一次性带上的,就得靠 Redis 一类内存存储补取。其二部署复杂——实时出预测令模型变更的滚动发布尤其困难,在 Kubernetes 这类容器编排环境更甚。其三可扩展性受限——模型会随时改变故不能简单水平扩展,可能要建单模型实例集群尽快消费新数据、并把学到的参数随 API 响应返回。其四监督要求更高——实时变化可能混入恶意行为并影响模型,需更先进的监控与调整/回滚机制。其五管理要求更高——做对需要数据科学与产品工程两方面的不菲经验与精调,因此大概只值得用于高金额影响的关键业务线;且与 MaaS 结合才更有力。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • Serving · 查看原书依据

      Disadvantages Here are some disadvantages of the online-serving approach: Latency budget required The model needs access to all relevant features. It will need quick access to new queries so that it can convert them into features and in turn look up relevant features stored elsewhere. If all the data we need for a single training example can’t be sent to the server as part of the payload on the API call, we need to grab that d

    参考来源

    - 位置:《Reliable Machine Learning》第08章《Serving》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch08.xhtml) - 结论依据:原文明确阐述该主题,并给出定义、实践或反例。 - 原始内容:Disadvantages Here are some disadvantages of the online-serving approach: Latency budget required The model needs access to all relevant features. It will need quick access to new queries so that it can convert them into features and in turn look up relevant features stored elsewhere. If all the data we need for a single training example can’t be sent to the server as part of the payload on the API call, we need to grab that d