{"type":"doc","content":[{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"前言"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在"},{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/u\/banyu\/publish","title":"xxx","type":null},"content":[{"type":"text","text":"伴鱼"}]},{"type":"text","text":"，我们在多个在线场景使用机器学习提高用户的使用体验，例如：在伴鱼绘本中，我们根据用户的帖子浏览记录，为用户推荐他们感兴趣的帖子；在转化后台里，我们根据用户的绘本购买记录，为用户推荐他们可能感兴趣的课程等。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征是机器学习模型的输入。如何高效地将特征从数据源加工出来，让它能够被在线服务高效地访问，决定了我们能否在生产环境可靠地使用机器学习。为此，我们搭建了特征系统，系统性地解决这一问题。目前，伴鱼的机器学习特征系统运行了接近 100 个特征，支持了多个业务线的模型对在线获取特征的需求。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"下面，我们将介绍特征系统在伴鱼的演进过程，以及其中的权衡考量。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"特征系统 V1"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征系统 V1 由三个核心组件构成：特征管道，特征仓库，和特征服务。整体架构如下图所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/21\/c5\/2194e3e2088f0317fb62cd4c12dec4c5.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征管道包括流特征管道和批特征管道，它们分别消费流数据源和批数据源，对数据经过预处理加工成特征（这一步称为特征工程），并将特征写入特征仓库。批特征管道使用 "},{"type":"link","attrs":{"href":"https:\/\/spark.apache.org\/","title":"xxx","type":null},"content":[{"type":"text","text":"Spark "}]},{"type":"text","text":"实现，由 DolphinScheduler 进行调度，跑在 YARN 集群上。出于技术栈的一致考虑，流特征管道使用 Spark Structured Streaming 实现，和批特征管道一样跑在 YARN 集群上。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征仓库选用合适的存储组件（Redis）和数据结构（Hashes），为模型服务提供低延迟的特征访问能力。之所以选用 Redis 作为存储，是因为："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"伴鱼有丰富的 Redis 使用经验。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"包括 "},{"type":"link","attrs":{"href":"https:\/\/doordash.engineering\/2020\/11\/19\/building-a-gigascale-ml-feature-store-with-redis\/","title":null,"type":null},"content":[{"type":"text","text":"DoorDash Feature Store"}]},{"type":"text","text":" 和 "},{"type":"link","attrs":{"href":"https:\/\/docs.feast.dev\/feast-on-kubernetes\/concepts\/stores#online-store","title":null,"type":null},"content":[{"type":"text","text":"Feast"}]},{"type":"text","text":" 在内的业界特征仓库解决方案都使用了 Redis。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征服务屏蔽特征仓库的存储和数据结构，对外暴露 RPC 接口 "},{"type":"codeinline","content":[{"type":"text","text":"GetFeatures(EntityName, FeatureNames)"}]},{"type":"text","text":"，提供对特征的低延迟点查询。在实现上，这一接口基本对应于 Redis 的 "},{"type":"codeinline","content":[{"type":"text","text":"HMGET EntityName FeatureName_1 ... FeatureName_N"}]},{"type":"text","text":" 操作。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这一版本的特征系统存在几个问题："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"算法工程师缺少控制，导致迭代效率低。这个问题与系统涉及的技术栈和公司的组织架构有关。在整个系统中，特征管道的迭代需求最高，一旦模型对特征有新的需求，就需要修改或者编写一个新的 Spark 任务。而 Spark 任务的编写需要有一定的 Java 或 Scala 知识，不属于算法工程师的常见技能，因此交由大数据团队全权负责。大数据团队同时负责多项数据需求，往往有很多排期任务。结果便是新特征的上线涉及频繁地跨部门沟通，迭代效率低。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征管道只完成了轻量的特征工程，降低在线推理的效率。由于特征管道由大数据工程师而非算法工程师编写，复杂的数据预处理涉及更高的沟通成本，因此这些特征的预处理程度都比较轻量，更多的预处理被留到模型服务甚至模型内部进行，增大了模型推理的时延。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"为了解决这几个问题，特征系统 V2 提出几个设计目的："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将控制权交还算法工程师，提高迭代效率。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将更高权重的特征工程交给特征管道，提高在线推理的效率。"}]}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"特征系统 V2"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征系统 V2 相比特征系统 V1 在架构上的唯一不同点在于，它将特征管道切分为三部分：特征生成管道，特征源，和特征注入管道。值得一提的是，管道在实现上均从 Spark 转为 Flink，和公司数据基础架构的发展保持一致。特征系统 V2 的整体架构如下图所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/71\/71\/71bb73d26679a34950dac22f2a16a071.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征生成管道读取原始数据源，加工为特征，并将特征写入指定特征源（而非特征仓库）。如果管道以流数据源作为原始数据源，则它是流特征生成管道；如果管道以批数据源作为原始数据源，则它是批特征生成管道。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征生成管道的逻辑由算法工程师全权负责编写。其中，批特征生成管道使用 HiveQL 编写，由 DolphinScheduler 调度。流特征生成管道使用 PyFlink 实现，详情见下图。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/dd\/50\/ddd1d9531a905b142eb5b23b24591950.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"算法工程师需要遵守下面步骤："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":null,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"用 Flink SQL 声明 Flink 任务源（source.sql）和定义特征工程逻辑（transform.sql）。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"（可选）用 Python 实现特征工程逻辑中可能包含的 UDF 实现（udf_def.py）。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"使用自研的代码生成工具，生成可执行的 PyFlink 任务脚本（run.py）。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":4,"align":null,"origin":null},"content":[{"type":"text","text":"本地使用由平台准备好的 Docker 环境调试 PyFlink 脚本，确保能在本地正常运行。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":5,"align":null,"origin":null},"content":[{"type":"text","text":"把代码提交到一个统一管理特征管道的代码仓库，由 AI 平台团队进行代码审核。审核通过的脚本会被部署到伴鱼实时计算平台，完成特征生成管道的上线。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这一套流程确保了："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"算法工程师掌握上线特征的自主权。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平台工程师把控特征生成管道的代码质量，并在必要时可以对它们实现重构，而无需算法工程师的介入。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征源存储从原始数据源加工形成的特征。值得强调的是，它同时还是连接算法工程师和 AI 平台工程师的桥梁。算法工程师只负责实现特征工程的逻辑，将原始数据加工为特征，写入特征源，剩下的事情就交给 AI 平台。平台工程师实现特征注入管道，将特征写入特征仓库，以特征服务的形式对外提供数据访问服务。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征注入管道将特征从特征源读出，写入特征仓库。由于 Flink 社区缺少对 Redis sink 的原生支持，我们通过拓展 "},{"type":"link","attrs":{"href":"https:\/\/github.com\/apache\/flink\/blob\/master\/flink-streaming-java\/src\/main\/java\/org\/apache\/flink\/streaming\/api\/functions\/sink\/RichSinkFunction.java","title":null,"type":null},"content":[{"type":"text","text":"RichSinkFunction"}]},{"type":"text","text":" 简单地实现了 "},{"type":"codeinline","content":[{"type":"text","text":"StreamRedisSink"}]},{"type":"text","text":" 和 "},{"type":"codeinline","content":[{"type":"text","text":"BatchRedisSink"}]},{"type":"text","text":"，很好地满足我们的需求。其中，"},{"type":"codeinline","content":[{"type":"text","text":"BatchRedisSink"}]},{"type":"text","text":" 实现了批量写入，大幅减少对 Redis server 的请求量，增大吞吐，将写入效率提升了 7 倍，见"},{"type":"link","attrs":{"href":"https:\/\/tech.ipalfish.com\/blog\/2021\/06\/25\/flink-bulk-insert-redis\/","title":null,"type":null},"content":[{"type":"text","text":"博客"}]},{"type":"text","text":"。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征系统 V2 很好地满足了我们提出的设计目的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"由于特征生成管道的编写只需用到 SQL 和 Python 这两种算法工程师十分熟悉的工具，因此他们全权负责特征生成管道的编写和上线，无需依赖大数据团队，大幅提高了迭代效率。在熟悉后，算法工程师通常只需花费半个小时以内，就可以完成流特征的编写、调试和上线。而这个过程原本需要花费数天，取决于大数据团队的排期。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"出于同样的原因，算法工程师可以在有需要的前提下，完成更重度的特征工程，从而减少模型服务和模型的负担，提高模型在线推理效率。"}]}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"总结"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征系统 V1 解决了特征上线的问题，而特征系统 V2 在此基础上，解决了特征上线难的问题。在特征系统的演进过程中，我们总结出作为平台研发的几点经验："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平台应该提供用户想用的工具。这与 Uber ML 平台团队在内部推广的"},{"type":"link","attrs":{"href":"https:\/\/eng.uber.com\/scaling-michelangelo\/","title":null,"type":null},"content":[{"type":"text","text":"经验"}]},{"type":"text","text":"相符。算法工程师在 Python 和 SQL 环境下工作效率最高，而不熟悉 Java 和 Scala。那么，想让算法工程师自主编写特征管道，平台应该支持算法工程师使用 Python 和 SQL 编写特征管道，而不是让算法工程师去学 Java 和 Scala，或是把工作转手给大数据团队去做。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平台应该提供易用的本地调试工具。我们提供的 Docker 环境封装了 Kafka 和 Flink，让用户可以在本地快速调试 PyFlink 脚本，而无需等待管道部署到测试环境后再调试。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平台应该在鼓励用户自主使用的同时，通过自动化检查或代码审核等方式牢牢把控质量。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作者：陈易生"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文：https:\/\/tech.ipalfish.com\/blog\/2021\/07\/30\/palfish-feature-system\/"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文：机器学习特征系统在伴鱼的演进"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"来源：伴鱼技术博客"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"转载：著作权归作者所有。商业转载请联系作者获得授权，非商业转载请注明出处。"}]}]}
机器学习特征系统在伴鱼的演进
{"type":"doc","content":[{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"前言"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在"},{"type":"link","attrs":{"href":"https:\/\/www.infoq.cn\/u\/banyu\/publish","title":"xxx","type":null},"content":[{"type":"text","text":"伴鱼"}]},{"type":"text","text":"，我们在多个在线场景使用机器学习提高用户的使用体验，例如：在伴鱼绘本中，我们根据用户的帖子浏览记录，为用户推荐他们感兴趣的帖子；在转化后台里，我们根据用户的绘本购买记录，为用户推荐他们可能感兴趣的课程等。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征是机器学习模型的输入。如何高效地将特征从数据源加工出来，让它能够被在线服务高效地访问，决定了我们能否在生产环境可靠地使用机器学习。为此，我们搭建了特征系统，系统性地解决这一问题。目前，伴鱼的机器学习特征系统运行了接近 100 个特征，支持了多个业务线的模型对在线获取特征的需求。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"下面，我们将介绍特征系统在伴鱼的演进过程，以及其中的权衡考量。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"特征系统 V1"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征系统 V1 由三个核心组件构成：特征管道，特征仓库，和特征服务。整体架构如下图所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/21\/c5\/2194e3e2088f0317fb62cd4c12dec4c5.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征管道包括流特征管道和批特征管道，它们分别消费流数据源和批数据源，对数据经过预处理加工成特征（这一步称为特征工程），并将特征写入特征仓库。批特征管道使用 "},{"type":"link","attrs":{"href":"https:\/\/spark.apache.org\/","title":"xxx","type":null},"content":[{"type":"text","text":"Spark "}]},{"type":"text","text":"实现，由 DolphinScheduler 进行调度，跑在 YARN 集群上。出于技术栈的一致考虑，流特征管道使用 Spark Structured Streaming 实现，和批特征管道一样跑在 YARN 集群上。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征仓库选用合适的存储组件（Redis）和数据结构（Hashes），为模型服务提供低延迟的特征访问能力。之所以选用 Redis 作为存储，是因为："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"伴鱼有丰富的 Redis 使用经验。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"包括 "},{"type":"link","attrs":{"href":"https:\/\/doordash.engineering\/2020\/11\/19\/building-a-gigascale-ml-feature-store-with-redis\/","title":null,"type":null},"content":[{"type":"text","text":"DoorDash Feature Store"}]},{"type":"text","text":" 和 "},{"type":"link","attrs":{"href":"https:\/\/docs.feast.dev\/feast-on-kubernetes\/concepts\/stores#online-store","title":null,"type":null},"content":[{"type":"text","text":"Feast"}]},{"type":"text","text":" 在内的业界特征仓库解决方案都使用了 Redis。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征服务屏蔽特征仓库的存储和数据结构，对外暴露 RPC 接口 "},{"type":"codeinline","content":[{"type":"text","text":"GetFeatures(EntityName, FeatureNames)"}]},{"type":"text","text":"，提供对特征的低延迟点查询。在实现上，这一接口基本对应于 Redis 的 "},{"type":"codeinline","content":[{"type":"text","text":"HMGET EntityName FeatureName_1 ... FeatureName_N"}]},{"type":"text","text":" 操作。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这一版本的特征系统存在几个问题："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"算法工程师缺少控制，导致迭代效率低。这个问题与系统涉及的技术栈和公司的组织架构有关。在整个系统中，特征管道的迭代需求最高，一旦模型对特征有新的需求，就需要修改或者编写一个新的 Spark 任务。而 Spark 任务的编写需要有一定的 Java 或 Scala 知识，不属于算法工程师的常见技能，因此交由大数据团队全权负责。大数据团队同时负责多项数据需求，往往有很多排期任务。结果便是新特征的上线涉及频繁地跨部门沟通，迭代效率低。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征管道只完成了轻量的特征工程，降低在线推理的效率。由于特征管道由大数据工程师而非算法工程师编写，复杂的数据预处理涉及更高的沟通成本，因此这些特征的预处理程度都比较轻量，更多的预处理被留到模型服务甚至模型内部进行，增大了模型推理的时延。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"为了解决这几个问题，特征系统 V2 提出几个设计目的："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将控制权交还算法工程师，提高迭代效率。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"将更高权重的特征工程交给特征管道，提高在线推理的效率。"}]}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"特征系统 V2"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征系统 V2 相比特征系统 V1 在架构上的唯一不同点在于，它将特征管道切分为三部分：特征生成管道，特征源，和特征注入管道。值得一提的是，管道在实现上均从 Spark 转为 Flink，和公司数据基础架构的发展保持一致。特征系统 V2 的整体架构如下图所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/71\/71\/71bb73d26679a34950dac22f2a16a071.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征生成管道读取原始数据源，加工为特征，并将特征写入指定特征源（而非特征仓库）。如果管道以流数据源作为原始数据源，则它是流特征生成管道；如果管道以批数据源作为原始数据源，则它是批特征生成管道。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征生成管道的逻辑由算法工程师全权负责编写。其中，批特征生成管道使用 HiveQL 编写，由 DolphinScheduler 调度。流特征生成管道使用 PyFlink 实现，详情见下图。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/dd\/50\/ddd1d9531a905b142eb5b23b24591950.jpg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"算法工程师需要遵守下面步骤："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"numberedlist","attrs":{"start":null,"normalizeStart":1},"content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":1,"align":null,"origin":null},"content":[{"type":"text","text":"用 Flink SQL 声明 Flink 任务源（source.sql）和定义特征工程逻辑（transform.sql）。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":2,"align":null,"origin":null},"content":[{"type":"text","text":"（可选）用 Python 实现特征工程逻辑中可能包含的 UDF 实现（udf_def.py）。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":3,"align":null,"origin":null},"content":[{"type":"text","text":"使用自研的代码生成工具，生成可执行的 PyFlink 任务脚本（run.py）。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":4,"align":null,"origin":null},"content":[{"type":"text","text":"本地使用由平台准备好的 Docker 环境调试 PyFlink 脚本，确保能在本地正常运行。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":5,"align":null,"origin":null},"content":[{"type":"text","text":"把代码提交到一个统一管理特征管道的代码仓库，由 AI 平台团队进行代码审核。审核通过的脚本会被部署到伴鱼实时计算平台，完成特征生成管道的上线。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这一套流程确保了："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"算法工程师掌握上线特征的自主权。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平台工程师把控特征生成管道的代码质量，并在必要时可以对它们实现重构，而无需算法工程师的介入。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征源存储从原始数据源加工形成的特征。值得强调的是，它同时还是连接算法工程师和 AI 平台工程师的桥梁。算法工程师只负责实现特征工程的逻辑，将原始数据加工为特征，写入特征源，剩下的事情就交给 AI 平台。平台工程师实现特征注入管道，将特征写入特征仓库，以特征服务的形式对外提供数据访问服务。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征注入管道将特征从特征源读出，写入特征仓库。由于 Flink 社区缺少对 Redis sink 的原生支持，我们通过拓展 "},{"type":"link","attrs":{"href":"https:\/\/github.com\/apache\/flink\/blob\/master\/flink-streaming-java\/src\/main\/java\/org\/apache\/flink\/streaming\/api\/functions\/sink\/RichSinkFunction.java","title":null,"type":null},"content":[{"type":"text","text":"RichSinkFunction"}]},{"type":"text","text":" 简单地实现了 "},{"type":"codeinline","content":[{"type":"text","text":"StreamRedisSink"}]},{"type":"text","text":" 和 "},{"type":"codeinline","content":[{"type":"text","text":"BatchRedisSink"}]},{"type":"text","text":"，很好地满足我们的需求。其中，"},{"type":"codeinline","content":[{"type":"text","text":"BatchRedisSink"}]},{"type":"text","text":" 实现了批量写入，大幅减少对 Redis server 的请求量，增大吞吐，将写入效率提升了 7 倍，见"},{"type":"link","attrs":{"href":"https:\/\/tech.ipalfish.com\/blog\/2021\/06\/25\/flink-bulk-insert-redis\/","title":null,"type":null},"content":[{"type":"text","text":"博客"}]},{"type":"text","text":"。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征系统 V2 很好地满足了我们提出的设计目的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"由于特征生成管道的编写只需用到 SQL 和 Python 这两种算法工程师十分熟悉的工具，因此他们全权负责特征生成管道的编写和上线，无需依赖大数据团队，大幅提高了迭代效率。在熟悉后，算法工程师通常只需花费半个小时以内，就可以完成流特征的编写、调试和上线。而这个过程原本需要花费数天，取决于大数据团队的排期。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"出于同样的原因，算法工程师可以在有需要的前提下，完成更重度的特征工程，从而减少模型服务和模型的负担，提高模型在线推理效率。"}]}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"总结"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特征系统 V1 解决了特征上线的问题，而特征系统 V2 在此基础上，解决了特征上线难的问题。在特征系统的演进过程中，我们总结出作为平台研发的几点经验："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平台应该提供用户想用的工具。这与 Uber ML 平台团队在内部推广的"},{"type":"link","attrs":{"href":"https:\/\/eng.uber.com\/scaling-michelangelo\/","title":null,"type":null},"content":[{"type":"text","text":"经验"}]},{"type":"text","text":"相符。算法工程师在 Python 和 SQL 环境下工作效率最高，而不熟悉 Java 和 Scala。那么，想让算法工程师自主编写特征管道，平台应该支持算法工程师使用 Python 和 SQL 编写特征管道，而不是让算法工程师去学 Java 和 Scala，或是把工作转手给大数据团队去做。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平台应该提供易用的本地调试工具。我们提供的 Docker 环境封装了 Kafka 和 Flink，让用户可以在本地快速调试 PyFlink 脚本，而无需等待管道部署到测试环境后再调试。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"平台应该在鼓励用户自主使用的同时，通过自动化检查或代码审核等方式牢牢把控质量。"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"作者：陈易生"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文：https:\/\/tech.ipalfish.com\/blog\/2021\/07\/30\/palfish-feature-system\/"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文：机器学习特征系统在伴鱼的演进"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"来源：伴鱼技术博客"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"转载：著作权归作者所有。商业转载请联系作者获得授权，非商业转载请注明出处。"}]}]}
發表評論
所有評論
還沒有人評論，想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章