深入解析 Flink 的算子链机制

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"“为什么我的 Flink 作业 Web UI 中只显示出了一个框，并且 Records Sent 和Records Received 指标都是 0 ？是我的程序写得有问题吗？”"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"Flink 算子链简介"}]},{"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":"笔者在 Flink 社区群里经常能看到类似这样的疑问。这种情况几乎都不是程序有问题，而是因为 Flink 的 operator chain ——即算子链机制导致的，即提交的作业的执行计划中，所有算子的并发实例（即 sub-task ）都因为满足特定条件而串成了整体来执行，自然就观察不到算子之间的数据流量了。当然上述是一种特殊情况。我们更常见到的是只有部分算子得到了算子链机制的优化，如官方文档中出现过多次的下图所示，注意 Source 和 map() 算子。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/72/720b4e0e2709778ba105ae601308d74d.png","alt":null,"title":"","style":[{"key":"width","value":"100%"},{"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}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"算子链机制的好处是显而易见的：所有 chain 在一起的 sub-task 都会在同一个线程（即 TaskManager 的 slot）中执行，能够减少不必要的数据交换、序列化和上下文切换，从而提高作业的执行效率。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/4b/4b9122f479e9ddeb5b03394bca2d367a.png","alt":null,"title":"","style":[{"key":"width","value":"100%"},{"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}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"铺垫了这么多，接下来就通过源码简单看看算子链产生的条件，以及它是如何在 Flink Runtime 中实现的。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"逻辑计划中的算子链"}]},{"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":"对 Flink Runtime 稍有了解的看官应该知道，Flink 作业的执行计划会用三层图结构来表示，即："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" StreamGraph —— 原始逻辑执行计划"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" JobGraph —— 优化的逻辑执行计划（Web UI 中看到的就是这个）"}]}]},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":" ExecutionGraph —— 物理执行计划"}]}]}]},{"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":"算子链是在优化逻辑计划时加入的，也就是由 StreamGraph 生成 JobGraph 的过程中。那么我们来到负责生成 JobGraph 的 o.a.f.streaming.api.graph.StreamingJobGraphGenerator 类，查看其核心方法 createJobGraph() 的源码。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":null},"content":[{"type":"text","text":"private JobGraph createJobGraph() {\n // make sure that all vertices start immediately\n jobGraph.setScheduleMode(streamGraph.getScheduleMode());\n // Generate deterministic hashes for the nodes in order to identify them across\n // submission iff they didn't change.\n Map hashes = defaultStreamGraphHasher.traverseStreamGraphAndGenerateHashes(streamGraph);\n // Generate legacy version hashes for backwards compatibility\n List