Twitter Storm中Bolt消息传递路径之源码解读

本文初次发表于storm-cn的google groups中，现以blog的方式再次发表，表明本人徽沪一郎确实读过这些代码，:).

Bolt作为task被executor执行，而executor是一个个的线程，所以executor必须存在于具体的process之中，而这个process就是worker。至于worker是如何被supervisor创建，尔后worker又如何创建executor线程，这些暂且按下不表。

 

假设同属于一个Topology的Spout与Bolt分别处于不同的JVM，即不同的worker中，不同的JVM可能处于同一台物理机器，也可能处于不同的物理机器中。为了让情景简单，认为JVM处于不同的物理机器中。

 

Spout的输出消息到达Bolt，作为Bolt的输入会经过这么几个阶段。

 

1. spout的输出通过该spout所处worker的消息输出线程，将tuple输入到Bolt所属的worker。它们之间的通路是socket连接，用ZeroMQ实现。

2. bolt所处的worker有一个专门处理socket消息的receive thread 接收到spout发送来的tuple

3. receive thread将接收到的消息传送给对应的bolt所在的executor。 在worker内部（即同一process内部),消息传递使用的是Lmax Disruptor pattern.

4. executor接收到tuple之后，由event-handler进行处理

 

下面是具体的源码

1. worker创建消息接收线程 

 

worker.clj

 

(defn launch-receive-thread [worker]

  (log-message "Launching receive-thread for " (:assignment-id worker) ":" (:port worker))

  (msg-loader/launch-receive-thread!

    (:mq-context worker)

    (:storm-id worker)

    (:port worker)

    (:transfer-local-fn worker)

    (-> worker :storm-conf (get TOPOLOGY-RECEIVER-BUFFER-SIZE))

    :kill-fn (fn [t] (halt-process! 11))))

 

注意加亮的行会将storm.yaml中配置使用ZMQ或其它

storm.messaging.transport:"backtype.storm.messaging.zmq"

 

2. worker从socket接收到新消息

vthread (async-loop

                 (fn []

                   (let [socket (.bind ^IContext context storm-id port)]

                     (fn []

                       (let [batched (ArrayList.)

                             init (.recv ^IConnection socket 0)]

                         (loop [packet init]

                           (let [task (if packet (.task ^TaskMessage packet))

                                 message (if packet (.message ^TaskMessage packet))]

                             (if (= task -1)

                               (do (log-message "Receiving-thread:[" storm-id ", " port "] received shutdown notice")

                                 (.close socket)

                                 nil )

                               (do

                                 (when packet (.add batched [task message]))

                                 (if (and packet (< (.size batched) max-buffer-size))

                                   (recur (.recv ^IConnection socket 1))

                                   (do (transfer-local-fn batched)

                                     0 ))))))))))

 

加亮行使用的transfer-local-fn会将接收的TaskMessage传递给相应的executor

 

3. transfer-local-fn

 

(defn mk-transfer-local-fn [worker]

  (let [short-executor-receive-queue-map (:short-executor-receive-queue-map worker)

        task->short-executor (:task->short-executor worker)

        task-getter (comp #(get task->short-executor %) fast-first)]

    (fn [tuple-batch]

      (let [grouped (fast-group-by task-getter tuple-batch)]

        (fast-map-iter [[short-executor pairs] grouped]

          (let [q (short-executor-receive-queue-map short-executor)]

            (if q

              (disruptor/publish q pairs)

              (log-warn "Received invalid messages for unknown tasks. Dropping... ")

              )))))))

 

用disruptor在线程之间进行消息传递。

 

多费一句话，mk-transfer-local-fn表示将外部世界的消息传递给本进程内的线程。而mk-transfer-fn则刚好在方向上反过来。

 

4. 消息被executor处理

 

executor.clj

==========================================================

(defn mk-task-receiver [executor-data tuple-action-fn]

  (let [^KryoTupleDeserializer deserializer (:deserializer executor-data)

        task-ids (:task-ids executor-data)

        debug? (= true (-> executor-data :storm-conf (get TOPOLOGY-DEBUG)))

        ]

    (disruptor/clojure-handler

      (fn [tuple-batch sequence-id end-of-batch?]

        (fast-list-iter [[task-id msg] tuple-batch]

          (let [^TupleImpl tuple (if (instance? Tuple msg) msg (.deserialize deserializer msg))]

            (when debug? (log-message "Processing received message " tuple))

            (if task-id

              (tuple-action-fn task-id tuple)

              ;; null task ids are broadcast tuples

              (fast-list-iter [task-id task-ids]

                (tuple-action-fn task-id tuple)

                ))

            ))))))

 

加亮行中tuple-action-fn定义于mk-threads(源文件executor.clj)中。因为当前以Bolt为例，所以会调用的tuple-action-fn定义于defmethod mk-threads :bolt [executor-data task-datas]

 

那么mk-task-receiver是如何与disruptor关联起来的呢，可以见定义于mk-threads中的下述代码

(let [receive-queue (:receive-queue executor-data)

              event-handler (mk-task-receiver executor-data tuple-action-fn)]

          (disruptor/consumer-started! receive-queue)

          (fn []            

            (disruptor/consume-batch-when-available receive-queue event-handler)

            0)))

 

到了这里，消息的发送与接收处理路径打通。