字节跳动万亿级图数据库的应用与挑战

{"type":"doc","content":[{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"一、了解图数据库"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1、开胃菜：公司业务场景的难题"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/c5\/c586636f4c66fd99e01b53c666769aff.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"首先介绍字节跳动的一个产品——抖音，作为我们本次分享的开胃菜。大家都知道目前抖音的视频的推荐，是通过算法去做的，这当中肯定会有一些最核心的基础数据的存储，比如说用户之间的关注关系、视频点赞的这种关系，以及用户的相互之间的通讯录的关系，是基于这些关系的数据去做推荐算法的。"}]},{"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":"如何去存储这些数据呢？这本身是一个非常有挑战性的问题，所以基于用户的关系来做这个内容或者是用户的推荐，从算法的角度上来说，是一个核心指标提升很大的事情。"}]},{"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":"做推荐显然是会基于二度和多度的关系，基于pattern来做推荐的，pattern本身是有很多种组合的，如何去基于这些组合去求解，这本身也是作为技术架构需要有这样的能力。"}]},{"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":"那么我们现在先假设一下原来的时限。比如说我要做二度关系，根据二度邻居的关注关系的结果做查询，那么以前的方法是首先要从线上MySQL，把用户的关系dump到hive上，然后将多张的hive表做这样的一个join，然后产出这样的二度关系，然后再将二度关系导入一个在线的KV系统，用于做推荐，然后上述的过程每天会去执行几次。"}]},{"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":"那可以想象在6亿+DAU的抖音，它的整个数据量是非常大的，所以这样的一条条数据下来，它的时间成本和算力成本也是非常大的。"}]},{"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":"因此，通过算法推荐这个方式，得到的数据已经很旧了，可能是几天以前的数据，所以导致它推荐不实时，以及策略迭代的代价太大了，中间任何一个环节出问题，可能就会导致整个过程的失败。"}]},{"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":"第三个问题就是日活如此高的抖音，如何实现高并发，以及毫秒级的查询延时，是比较有挑战的一件事情。"}]},{"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，我们来看看图数据库的好处是什么？"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2、图数据库与关系型数据库"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"我们需要做一个简单的分析，图数据库相比于传统的关系型数据库，它的区别是什么？"}]},{"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":"首先图非常很简单，由最基础的点、边以及属性所构成的，而关系型数据库当中，当我对基于一种关系或者说某种特定的条件去做查询的时候，通常的一个技术是多表的join。"}]},{"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":"那图数据库对于同样的工作量，它的反应是通过这个图上的遍历（traversal）去实现的，操作更加高效。"}]},{"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":"因为图数据库本身是基于点和边所组成的，所以从固定的点去做traversal，它的形成消息分发是一个更加细粒度的分发，所以相当于在一个大的分布式架构上，它会变得更细粒度，对内存的开销、网络的开销都会变得更小，这是背后的系统层面。"}]},{"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":"简单举一个例子：我如何去做一鸣的好友，然后我再去查询好友的公司有多少名员工？"}]},{"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":"在关系型数据库当中通常去需要有这样的一些 table，比如说公司的信息表、雇佣关系表、员工信息表等，用MySQL去查，可能就需要写成如下图这样的 SQL语句。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/41\/415d0477c21f72cbd1c68f62cbc70dc3.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"而基于图数据库，我们用gremlin这样一种查询语言去实现，简单的一行就能把它非常直白地写完。"}]},{"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":"接下来我会简单介绍一下图数据库里面流行的语言，以及整个图数据库里面的体系。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"3、图数据库业界对角度分类"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从语言生态、架构设计、集群规模、场景上分图数据库："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/c5\/c57774a2c65f2f73355b3613b15dfa0e.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"二、适用场景介绍举例"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1、ByteGraph的发展历程和适用业务数据模型"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"接下来，给大家介绍一下字节跳动内部ByteGraph的发展历程："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/d6\/d6af2ce9347d99c1652bf628215888b0.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"当前版本的ByteGraph在单机上是拥有百万级QPS的查询性能，并且同时支持了多维度的排序，比如说按照关注时间或者关系的亲密度，分别去做排序，也就是分别会基于时间和关系的亲密度去构建索引。"}]},{"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":"同时我们也支持会比较支持比较复杂的这种查询，比如说多跳的traversal，如果我要基于某一个点做一跳进而做二跳、三跳……通常这样的查询所涉及到的点的数量是非常多的，所以通常中间的某些子查询是可以并发去做的，来降低延时。"}]},{"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":"查询语言上我们当前是基于Gremlin去做的，Gremlin本身是一个图灵完备的语言，能表达任意的查询；作为一个公司百万级数据的这样的一个产品，可靠性是我们非常强调的，我们支持多机房容灾，然后也支持数据的最终一致性。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2、已上线业务场景分类"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"从业务上划分，目前我们支持了超过500多个业务集群，服务器规模已经达到上万台服务器。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/40\/40e4e03216b9b90c23a70ea0408897ee.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"举一个例子：最开始的一个业务，抖音去存储用户关系的在线存储，比如说好友关系、粉丝列表等，也有基于这些基础数据去做推荐的，比如说抖音推荐、推人、推视频等，是基于好友的好友等多跳的查询去做挖掘关系，然后做关联规则分析等一些算法上的内容。"}]},{"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":"另外也有知识图谱领域，支持搜索百科、教育团队、电商团队等，然后去做个体的推荐。另外的，IT系统上去用graph来抽象 rapo的依赖关系，或者线上服务之间的网络状态等。"}]},{"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":"这里举个例子，比如说抖音电商把它建模成一张图，首先实体（点）和联系（边）具体可以分成如下："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/fe\/feefb0a4c663ae4c2f9242d0086cfebc.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"我们会基于不同的设备、商品、达人等构建出来非常大的一张构图，然后这张图上会做各种类型的推荐，或者是离线、在线的分析，甚至是在线的基于图神经网络的训练等，有各种各样的应用。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"三、数据模型和查询语言"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1、有向属图建模"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"ByteGraph是基于有向属性图来建模的，有向属性图就是有点、边和属性来构成。点来表示的一些实体，然后通常ByteGraph内部是有一个二元组来唯一标志一个点，二元组通常有一个用户的 uid和他的在具体的一个应用场景下，比如说抖音、火山、头条等应用，用不同的应用分类成不同的垂直的场景，用这样的一个二元组来唯一标记一个实体点，然后 type标记的是一类点的集合。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/05\/05baf2a654b3e5e677e7816b018fa7e5.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"比如说一个运动员或者一个team，他表达的是真实世界当中的一个实体，可以简单理解成把table理解成是关系数据当中一个table，不同的 type是不同的table。"}]},{"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":"属性其实就是用来描述固定的一个实体，比如说姓名、性别、年龄等等，然后我们会规定同一种类型的 type点（schema）必须是一致的。"}]},{"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":"同理，边也是有一个这样的实体，它是由点之间的关系就映射成了边，它通常由起点加终点，以及边上的边类型type来描述一个事件。然后同时也会有很多种属性，比如说它的起始时间、终止时间、发生的地点等，来描述这样的一个事件。"}]},{"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":"简单介绍完了一下ByteGraph的图建模，我们再简单介绍一下ByteGraph的使用查询语言。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2、Gremlin查询语言接口"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Gremlin语言我刚才说的是一种图灵完备的语言，它隶属于Apache，是Apache的一个项目之一，它规定了 Gremlin的一些不同算子所涉及到的查询语义，但是对具体实践是不会有一个硬性的限制的，所以这取决于不同的厂商对Gremlin的标准，完全依赖于每一个厂商自己的实现。"}]},{"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":"目前ByteGraph支持了Gremlin的一个子集，覆盖率到了80%左右。数据模型就是有效属性图模型，Gremlin相比于传统RPC接口更灵活、表达力更好。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/3c\/3c5e13508f11a05447ac00b28dcfa3dd.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"从上述例子也可以看出，Gremlin这个语言是非常接近于自然语言的。"}]},{"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":"简单总结一下，基于Gremlin的查询语言接口，用学英语来比喻的话，就分为：学单词、组句子、开口讲这三步。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/e5\/e5e011f8d4ce6dae2e62d4cd4477d832.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"再用另外一个非常具体的 UGC场景来举例，基于Gremlin语言如何去表达不同类型的查询。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/0d\/0dba2fd7f973ed66adc4df7514d5cb61.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"可以看到Gremlin这个查询的写法，相对来说比SQL还要简单一点且非常直白，比如说我要去限制关注的大v的一个条件，会写一个where，where里面会写 otherV，otherV表示当前基于关注关系的这样的一条边，对应的vertex一定要限制是一个点，所以这样的一些条件限制可以写到where语句，然后我们会基于这个时间和 tsUs然后去倒排，然后限制直取Top10，最后limit（10），然后我会把这些大v作者的名字给取出来，这个时候会再去拿到otherV然后去求它的一个value，然后 value的属性名是name。"}]},{"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":"同时ByteGraph的特点是我们支持跨集群以及跨表的查询，下面还是以 UGC场景举例。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/ff\/ffb0c5af9f40270b24a836cddd1076b2.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"假设当前最上游的垂直业务当中，用户的关系是通过一张table来存储的，而点赞关系图通过另外一个table来存的。现在我想这个去查询，比如说用户C的好友（相互关注）所喜欢的文章的列表，则是我是基于用户C从table2开始去找，去找在table1当中文章的名字，这个时候我需要通过with Tbale语义去限制在table2中去找vertex C,然后从具有了double关注的所有用户当中，切到table1把具有的点赞关系的文章数它给列出来。"}]},{"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":"所以Gremlin它也支持这样子的一些跨表查询的语义，来支撑跨表甚至是跨集群的查询。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/e7\/e7e6593e6e3a5ec1b03de275dcea74e8.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"目前ByteGraph支持了 C++\/go SDK，上右图是基于Gremlin的语言，如果要去做一个图查询的话，怎样基于Java或者是C++、Python这些SDK去写这样的一个基于Gremlin的查询的例子。"}]},{"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":"目前的话，4种语言都支持 RPC的接口，但是基于Gremlin是暂时只支持C++和go的接口。"}]},{"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":"接下开会更深入的讲解一下 ByteGraph的整体架构。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"四、ByteGraph架构与实现"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1、ByteGraph整体架构"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"整个系统可以分成三层：查询引擎层、存储引擎层，磁盘存储层。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/71\/71cb0c17cc56e26dc081b5a341c2f12e.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"三层是相互独立的，每一层都可以水平扩容，比如说你查询的语言语义非常复杂，可能涉及到的step数比较多，所以它是一个计算比较重的这样的查询，但=内存存储开销可能会比较少一点，所以可以开更多的查询引擎层（GQ）的实例，开更少的存储引擎层（GS）实例。"}]},{"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":"整体上ByteGraph是基于了一个经典的计算存储分离的架构去做的，最底层是一个分布式KV，分布式KV我们用的是公司内部的一个叫Abase的一个分布式KV系统，同时我们也支持其他的分布式KV，比如说公司内部其实有一个 byte KV这样的一个产品，它跟Abase区别是一个是一致性、一个重可靠性，所以整体上后端引擎是可以以热插拔的形式去做更替的。"}]},{"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":"查询引擎层"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"主要涉及到用户session管理、服务的proxy，然后核心的一个功能是基于用户发过来的Gremlin请求，去做一个逻辑的逻辑查询计划的这样的一个生成。然后基于逻辑上的查询计划，生成一个物理的查询计划，然后通过执行器executor把对应的子查询给分发出去，它是由go来实现的，重高并发问题。"}]},{"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":"存储引擎层"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这边主要涉及到数据的存储，因此我们会在这个模块当中会涉及到如何把数据做切片，然后分成一个Graph partition（Graph shard），然后如何去把不同的shard内部所代表的子图，用一种特定的数据结构把它组织起来，同时这个数据结构要有相对来说比较良好且较低的读写放大能力，以及它能够在磁盘的组织形式上对磁盘比较友好，然后是顺序读写。"}]},{"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":"所以如何去实现这样的数据结构，也是我们ByteGraph比较一个核心的设计。为了保证数据不能丢失，在存储引擎层我们也支持了 WAL（Write-Ahead Log），同时我们也支持事务性，通过1PC的事务协议来支持的，事务性目前是当前是支持 Read Committed的事务隔离级别，这一层是通过C++写的。"}]},{"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":"磁盘存储层"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"当前是依赖于公司的第三方permission store（KV store）去做的，下个版本会自研图原生存储。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2、ByteGraph读写流程"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"然后我们简单分析一下，一个读写的query进来之后，它的路由机制怎样的呢？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/b1\/b141f45b0f34ab31c241ecdb3505e17f.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"假设当前的 GQ和GS层分别有不同的实例，一个写语句进来之后，它会基于当前写的X然后假设根据最简单的哈希规则，它会不会映射到 GQ2的实例上，GQ收到read query之后，会基于路由规则把它打到 GS2上，然后GS2的cache层就会去找X存不存在。"}]},{"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":"如果不存在的话，会把当前page从存储的KV store里去把它给捞上来，然后把当前写过程给写进去，与此同时，我们会写入一条WAL这样的log，把它固化到KV store，防止数据更新的丢失。"}]},{"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":"如果是一个读的过程，就只是做查询，就相对来说更简单了，直接去基于一个GQ的实例去找应该在哪个GS的实例上，去找到 A这样的page，如果找不到就把它从磁盘捞上来，如果找得到就直接返回结果。"}]},{"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":"所以简单可以把 GS层理解成缓存层，但同时我们也不是简单的一个缓存，因为在这一层上我们也支持数据的事务性，还有数据的防丢失的能力。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"3、ByteGraph实现"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"1）查询引擎"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/cc\/ccf00b013cdff1827cea840e8c3d578a.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"这里再简单提及一下ByteGraph的查询引擎，接下来我会依次对查询引擎、存储引擎做详细的分享。查询引擎这边首先第一个是要做 parser，把一个string打进来之后，把它解析成一个查询的语法数，基于一定的优化规则，如 RBO（rule based optimization）和CBO（cost based optimization）去生成一个逻辑的查询计划。"}]},{"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":"通常query在垂直的业务上，所以query和pattern是比较相似的，所以我们为了防止查询计划多次的生成，所以一个查询计划基于一个模板的情况下，我们是有缓存的。"}]},{"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":"然后生成了查询计划之后，接下来的事情就是让 GQ层与GS层之间交互，能并行的查询尽量并行去做，不能做的话就只能串行的去查询，基于这样的一个依赖关系去串行的完成这样的查询。"}]},{"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":"同时GQ层，需要去理解在存储层上的分片逻辑，找到对应的一个数据，它在具体在GQ层还是在GS层上。"}]},{"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":"同时，一个带索引的查询，它在存储层上已经建了索引之后，这个查询显然不应该把它放到查询上去做，它应该放到存储上去做。所以我们这里也涉及到一些算子下推的一些优化。"}]},{"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":"在查询优化器上分成两类：第一种是基于规则的优化器，第二个是基于代价的优化。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"2）查询优化器"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/61\/616a6a7b436f9e6c13f1a5abf664c992.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"查询优化器分成两类：基于规则的优化和基于代价的优化。"}]},{"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":"在基于代价的优化中，可以用右图来表示。执行计划A非常显而易见的一个方式，就是做两票的 expand，我先找到他的一度邻居，然后依次让一度邻居去找到他的二度邻居，看有多少人当中是有他的。另外一个方式是找到了我的一跳领居之后，然后找到他的一跳入住邻居，然后依次去做一个join，那显然二会比一开销很多。"}]},{"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":"所以我们在用户写出这样的一个query之后，我们的优化器能够找到相对cost最低的一个查询优化的逻辑执行计划。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"3）图分区算法"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/d2\/d2485a998fca2c6fe0fac87993c3411f.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"图分区的话，这一块ByteGraph支持了不同策略的图分区方式，比如说最简单的基于点的起点，和边的类型进行一致性哈希的分区方式，目前的话是在大部分场景上都是基于这样的分区的算法来做的。"}]},{"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":"ByteGraph支持了不同策略的图分区方式，比如说最简单的基于点的起点，和边的类型进行一致性哈希的分区方式，目前的话是在大部分场景上都是基于这样的一个分区的算法来做的。"}]},{"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":"知识图谱的场景它的特点在于它的边类型是非常多的，所以删除之后映射到每一种类型的边的数量相对较少，小到单机是可以完全容纳这种类型的边的所有集合，所以就不要考虑点了，完全依据边的类型进行哈希分区。"}]},{"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":"这样的话在知识图谱这个场景下，它能大幅的降低查询中多度查询扇出的请求数量，也就是网络的开销，进而就可以降低了延时。"}]},{"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":"在社交场景当中，通常来说是一个点的度的分布，但是有一些点它的度特别大，有一些点它的度特别小，甚至没有人关注它。在这种情况下，我们是基于 Facebook16年的一篇论文，去实现了一个这样的，一个social hash这样的一个算法，来保证我们做多跳邻居查询的时候，它的网络开销是比较小的。"}]},{"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":"所以这种情况下， ByteGraph会优先让整个图导入之后去做一个离线的图分区算法，然后做完了之后再把对应的点和边，基于这个算法映射到不同的数据。"}]},{"type":"heading","attrs":{"align":null,"level":4},"content":[{"type":"text","text":"4）ByteGraph存储引擎实现"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/74\/748f25b83f98b02d820ac156fcd49618.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"接下来再讲一下存储引擎的一些细节，整体上说存储引擎这边可以把整个系统组件划分成这样的几层。"}]},{"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":"最上层是一个跟图语有关的读写的接口，然后中间这一层是涉及到如何去支持数据的事务性，以及我们如何把一个数据映射成一个图原生的存储，它的数据layout是什么样子，我们在这一层把它给解决了，然后同时我们也支持 WAL，来保证数据的更新是能够持久化的，不会有任何数据的更新的丢失。"}]},{"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":"最下面一层就是基于KV store这样的接口，我们支持了不同类型的 KV store，比如说一些开源的HBase、RocksDB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","marks":[{"type":"strong"}],"text":" ① 存储结构（一）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"简单讲一下如何机遇KV系统能够构建一个图结构？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/a8\/a8f73f585619e81b9e400785872a796a.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"基于 KV的一个建模，最简单且直观的方式就是一个KV对一条边，同时它的写放大也非常小。"}]},{"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":"所以写放大就是当你想更新一条数据，这个数据可能是有一个字节数，比如说X，但是你实际上更新的这样一个数据块是Y，如果Y远大于X的话，就是写放大是非常大的，当前这样建模它的写放大是非常小的，因为它的粒度很细，但是你可以想象它去做查询，做一跳领域的查询的时候，它的性能是退化的程度是非常大的，因为它涉及到大量的随机读写，它数据的局部性就没有了。"}]},{"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":"如果用一个KV去保存一个起点的所有边，显然这个数据的局部性就会好，但是它的写放大就会变得很大。比如说你改了当前对应的一个点上的一条边，其实整个ege历史都要被更改，这是我们设计上的一个权衡，需要做一个折中。"}]},{"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":"具体是怎么做的，我们用一个类似于B树的结构来建模Graph，对某一个点来说，一个点同一个边type的所有的终点是一个存储单元，也就是说我们把一个起点ID、起点type和边type，基于它去group by，具有相同值的所有边集合，我们会认为它是逻辑上属于一个分区的。"}]},{"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":"如果这个分区依然涉及到很多点怎么办呢？我们会把它作为一个二级的拆分，所以因此会涉及到 b树的多层级。"}]},{"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":"假设一个点，它基于某一种关注关系，粉丝数是1000万，其实可以想象用一级的一个page去存肯定是不够的，我们会把它拆分成多page。"}]},{"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","marks":[{"type":"strong"}],"text":" ② 存储结构（二）"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/d3\/d34101c300e8f5805625bc901d3a3567.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"第一层的page就叫Meta page，它其实只是去简单记录了一个映射，这1000万个邻居当中，我们基于每2000为一片，每一片我们把它称作为一个Edge page ，每一个Edge page又存储了2000个Edge，所以用这样一个多级拆分的这样的方式去降低了读写放大的问题，同时起到了一个非常平衡的设计。"}]},{"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":"总结下来就是单起点和某一种固定的边类型组成了一个B树，然后B树的每一个节点是一个KV队，然后这里涉及到完整性上的话，我们会限制每一个B树的写者只能是唯一的，以防止并发的写入导致 B树逻辑上的破坏。"}]},{"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":"刚才说到写放大的问题，我们具体在当前 B树的建模上，依然其实会存在写放大的问题。"}]},{"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","marks":[{"type":"strong"}],"text":" ③ 日志管理"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/90\/90b68ab05d51c71f59a7d5849453e89c.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"我们是如何进一步去优化写放大的？比如说一个写请求，过来了之后，我们其实是只会去写 WAL的，当它在内存当中的某一个B树的page，当一个写请求进来，它确实映射到page上的数据了，显然内存中的数据是需要被更改的，但是磁盘上的数据这个时候是不需要更改的。"}]},{"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":"这个时候，它会写一条相对来说尺寸比较小的WAL，再把它固化到 KV store里，只有当这个数据再次被从磁盘上捞到内存里的时候，我们会把原有的磁盘上的旧数据apply到新的WAL，然后就生成了最新的数据，然后把它放到内存里，通过这样的方式来缓解写放大的问题。"}]},{"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":"然后同时如刚才所说，为了维持B树的完整性，每个必须是有且唯一的一个WAL日志流。"}]},{"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","marks":[{"type":"strong"}],"text":" ④ 缓存实现"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/30\/30ea231f19c56283a1e84d0498d8f27c.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"关于缓存的话，我们去实现了自己的一个高性能的LRU Cache。这个不难理解，作为一个数据库的话，你需要有一个相对来说比较泛化的能力，不同的垂直的应用场景，它所涉及到读写的比例以及读写的QPS也是不一样的。"}]},{"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":"所以我们LRU Cache是支持了不同的策略，基于不同的频率的读出，触发的阈值也不一样，比如说我们一台物理机的内存，比如说我用到了60%，这个时候想再往上走可能就比较危险了，这个时候我就开始触发LRU Cache的能力，把不会经常用到page，要下刷写到磁盘上。"}]},{"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":"当数据规模不变的时候、写请求的流量增大的情况下，缓存与存储分离的模式，它的一个优点就是可以快速的扩容，也就是把 GS这一层单独的去加大请求的个数，来提高我的缓存的能力，但存储层的QPS的整个的容量是不变的。"}]},{"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":"下图是ByteGraph的存储层的全貌，单机的内存引擎就会长这个样子。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/17\/17e2984d330e74625893c98a7322d24c.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"首先会把整个图数据模型成基于一个特定的点和它的边类型，会把它抽象成一个B树的数据架构。"}]},{"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":"如果随着读写流进来，特别是写流会把一些page更新掉它的数据，同时会写一个WAL，这个时候page会变脏，所以我们会用一个 dirty page的一个link list去记录脏数据，脏数据积累到一定程度后，我们要把脏数据下刷到磁盘上，然后同时我们会有维护WAL这样的一个log流，然后同时也会有这样的一个LRU Cache来保证一个我们的物理机的内存开销是在一个阈值之上的，有一个上界有一个下界。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"五、关键问题分析"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"1、ByteGraph关键问题之一：索引"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"索引我们目前是支持了全局索引和局部索引："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/c6\/c6d9d0ce9882adcdc8c53600d4eaedbe.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"局部索引"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"对于给定的起点和边类型，然后对边上的属性构建索引，比如说我基于用户的年龄索引，显然基于默认的属性，比如说我们当前的默认属性是基于时间去索引，边就会基于时间去做排序，如果基于Edge去索引的话，那会基于Edge去做排序，这是两个不同的B树的组织方式。"}]},{"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":"全局索引"}]}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"基于一个属性值，能查到当前在整个Gragh里面，具有特定属性值的所有点的ID，这个是全局索引的定义。然后这里就涉及到数据一致性的问题了，它本身是有分布式事务的能力，所以我们通过分布式事务能力来维护了数据与索引之间的一致性。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"2、ByteGraph关键问题之二：热点读写"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"举个例子：比如说一个大v正在直播的时候，可能有很多人进入了他的直播间，回到刚才那个例子，我们是通过一个图来去模拟用户与电商之间关系的，所以当有不同的用户进到这个商家的时候，其实你可以想象成在 Graph里面会有很多边被写进来了，很多人进入一个特定的商家的时候，就会造成热点的写问题，同样读也是一样的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/7c\/7c7beb76ee77cc435891d3d6ec2695ec.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"3、ByteGraph关键问题之三：离线在线数据流融合"}]},{"type":"image","attrs":{"src":"https:\/\/static001.geekbang.org\/infoq\/30\/30d5063a4fa660b3b4c5ba2721a86ca9.jpeg","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"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":"存量数据导入：ByteGraph目前对存量的数据导入，比如它有不同的数据源存入MySQL\/Hive\/Redis\/Hbase等，我们是通过这样的一个公司内部的平台MapReduce去 Bulkload到我们的ByteGraph里了，这是存量的离线的数据导入。"}]},{"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":"在线数据实时写入：在线实时的写入是通过线上的服务调我们的Gremlin的SDK，或者是RPC的 SDK去写入，或者也可以通过Kafka等这种消息队列在线写入到ByteGraph里面。"}]},{"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":"在线数据天级快照：ByteGraph也支持天级的数据快照，把一天的数据完整的放到hive里，然后用来给上游的业务同学做离线分析或离线训练等。"}]},{"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","marks":[{"type":"strong"}],"text":">>>>"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"Q&A"}]},{"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","marks":[{"type":"strong"}],"text":"Q1：查询语言Gremlin和GSQL差别大吗？GSQL会成为标准吗？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"A1："},{"type":"text","text":"据我所知，GSQL目前确实有一种趋势会变成标准，原因是它跟SQL长得很像。但是我个人认为，我觉得Gremlin对用户更友好，是比较贴近于自然语言的，然后它是一个比较类pipeline的这样一种语言，所以天生就比较适合做查询计划的优化。"}]},{"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":"因为目前关系型数据库还是主流，所以大家SQL会熟悉一点，所以会轻易的从SQL转到GSQL上。至于GSQL会不会成为标准，不影响当前的一个事实是Gremlin，已经被很多大厂基于查询语言，去做了一个这样的不同的数据库产品出来。"}]},{"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","marks":[{"type":"strong"}],"text":"Q2：支持HA吗？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"A2："},{"type":"text","text":"当前暂时是不支持的。"}]},{"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","marks":[{"type":"strong"}],"text":"Q3：请问支持一些基本的图计算操作吗？比如计算三角形个数triangle counting\/listing。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"A3："},{"type":"text","text":"目前是不支持的，我们有另外的一套系统去支持它。"}]},{"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","marks":[{"type":"strong"}],"text":"Q4：貌似DGraph更简单，问下，什么原因不选择DGraph？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"A4："},{"type":"text","text":"字节有自己特定的场景，我们不仅有国内的数据，还有国外的数据，数据量特别大，每秒钟要支持的QPS也特别高，所以目前开源的数据库都不能满足我们公司内部的需求。"}]},{"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","marks":[{"type":"strong"}],"text":"Q5：请问老师我们对超级节点的查询有处理吗？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"A5："},{"type":"text","text":"这个是有的，我们用B树来model这个图，假设设定阈值为2000，超出2000就会分裂成两个page，以此类推。像我们抖音上人民日报的粉丝数有1亿+，相当于一个超级节点，在ByteGraph的维护下，目前性能上都没有任何问题。"}]},{"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","marks":[{"type":"strong"}],"text":"Q6：图数据库也分为类似的OLTP和OLAP吗？还是主要应用于OLTP的场景？"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"A6："},{"type":"text","text":"目前OLTP和OLAP都是各有侧重的。"}]},{"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","marks":[{"type":"strong"}],"text":"嘉宾介绍："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"陈宏智博士，目前在字节跳动基础架构组担任高级研发工程师角色，博士毕业于香港中文大学计算机科学与工程系，本科毕业于华中科技大学计算机学院，在计算机系统及数据库领域发表顶会论文及期刊(e.g：EuroSys\/ SoCC\/SIGMOD\/KDD\/TPDS等)十余篇，研究方向为分布式系统、分布式计算、大规模图存储\/计算\/训练系统；港府博士奖学金获得者，微软研究院“明日之星”荣誉获得者。"}]},{"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":"本文转载自：dbaplus社群（ID：dbaplus）"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文链接："},{"type":"link","attrs":{"href":"https:\/\/mp.weixin.qq.com\/s\/OC5ZH7Ve2GGjLTc9g2eXVw","title":"xxx","type":null},"content":[{"type":"text","text":"字节跳动万亿级图数据库的应用与挑战"}]}]}]}