知识卡片
列导向存储
存储方式主要由访问模式驱动。若不同模型各取数据的不同特征子集,列导向存储(每特征一列)是常见架构:只取所需字段,不必每次读整行,适合各用不同特征子集的 ML 训练管线;行导向则适合都用几乎全部字段的相似应用。另一关键点是“数据几乎总被多次读取”——模型开发是迭代的,应假设开发者会半持续地重训同一模型、每次可能读不同子集,据此建存储。元数据则是被人理解“上一版模型怎么搭的”的路线图,常被低估。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 存储
原书出处 · 1 处
- Data Management Principles · 查看原书依据
Given this, a column-oriented storage scheme with one column per feature is a common design architecture, especially for models training on structured data.7 Most readers will be familiar with row-oriented storage, in which every fetch of data from the database retrieves all of the fields of a matching row. This is an appropriate architecture for collections of applications that all use most or all of the data—in other words,
参考来源
- 位置:《Reliable Machine Learning》第02章《Data Management Principles》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch02.xhtml)
- 结论依据:原文对比行列存储与多模型读取模式,并强调复用假设与元数据价值。
- 原始内容:Given this, a column-oriented storage scheme with one column per feature is a common design architecture, especially for models training on structured data.7 Most readers will be familiar with row-oriented storage, in which every fetch of data from the database retrieves all of the fields of a matching row. This is an appropriate architecture for collections of applications that all use most or all of the data—in other words,