码迷,mamicode.com
首页 > 其他好文 > 详细

storm启动supervisor源码分析-supervisor.clj

时间:2014-09-24 23:03:27      阅读:383      评论:0      收藏:0      [点我收藏+]

标签:des   style   http   color   io   os   使用   java   ar   

supervisor是storm集群重要组成部分,supervisor主要负责管理各个"工作节点"。supervisor与zookeeper进行通信,通过zookeeper的"watch机制"可以感知到是否有新的任务需要认领或哪些任务被重新分配。我们可以通用执行bin/storm supervisor >/dev/null 2>&1 &来启动supervisor。bin/storm是一个python脚本,在这个脚本中定义了一个supervisor函数:

supervisor函数

 

def supervisor(klass="backtype.storm.daemon.supervisor"):
   """Syntax: [storm supervisor]

   Launches the supervisor daemon. This command should be run
   under supervision with a tool like daemontools or monit.

   See Setting up a Storm cluster for more information.
   (https://github.com/nathanmarz/storm/wiki/Setting-up-a-Storm-cluster)
   """
   cppaths = [STORM_DIR + "/log4j", STORM_DIR + "/conf"]
   jvmopts = parse_args(confvalue("supervisor.childopts", cppaths)) + [
       "-Dlogfile.name=supervisor.log",
       "-Dlog4j.configuration=storm.log.properties",
   ]
   exec_storm_class(
       klass,
       jvmtype="-server",
       extrajars=cppaths,
       jvmopts=jvmopts)

klass参数的默认值为backtype.storm.daemon.supervisor,backtype.storm.daemon.supervisor标识一个java类。STORM_DIR标识storm的安装目录,cppaths集合存放了log4j配置文件路径和storm配置文件storm.yaml路径,jvmopts存放传递给jvm的参数,包括log4j配文件路径、storm.yaml路径、log4j日志名称和log4j配置文件名称。exec_storm_class函数的逻辑比较简单,具体实现如下:

exec_storm_class函数

 

def exec_storm_class(klass, jvmtype="-server", jvmopts=[], extrajars=[], args=[], fork=False):  
   global CONFFILE  
   all_args = [  
       "java", jvmtype, get_config_opts(),  
       "-Dstorm.home=" + STORM_DIR,  
       "-Djava.library.path=" + confvalue("java.library.path", extrajars),  
       "-Dstorm.conf.file=" + CONFFILE,  
       "-cp", get_classpath(extrajars),  
   ] + jvmopts + [klass] + list(args)  
   print "Running: " + " ".join(all_args)  
   if fork:  
       os.spawnvp(os.P_WAIT, "java", all_args)  
   else:  
       os.execvp("java", all_args) # replaces the current process and never returns

get_config_opts()获取jvm的默认配置信息,confvalue("java.library.path", extrajars)获取storm使用的本地库JZMQ加载路径,get_classpath(extrajars)获取所有依赖jar包的完整路径,然后拼接一个java -cp命令运行klass的main方法。klass默认值为backtype.storm.daemon.supervisor,所以exec_storm_class函数最终调用backtype.storm.daemon.supervisor类的main方法。
backtype.storm.daemon.supervisor类定义在supervisor.clj文件中,定义如下:

backtype.storm.daemon.supervisor类

 

(ns backtype.storm.daemon.supervisor
 (:import [backtype.storm.scheduler ISupervisor])
 (:use [backtype.storm bootstrap])
 (:use [backtype.storm.daemon common])
 (:require [backtype.storm.daemon [worker :as worker]])
 (:gen-class
   :methods [^{:static true} [launch [backtype.storm.scheduler.ISupervisor] void]]))

       (bootstrap)
       ;; ... ...
   ;; 其他方法
   ;; ... ...
   (defn -main []
         (-launch (standalone-supervisor)))

:gen-class指示Clojure生成Java类backtype.storm.daemon.supervisor,并且声明一个静态方法launch,launch方法接收一个实现backtype.storm.scheduler.ISupervisor接口的实例作为参数。launch函数的参数是由standalone-supervisor函数生成的。standalone-supervisor函数定义如下:返回一个实现ISupervisor接口的实例。

standalone-supervisor函数

 

;; 该函数主要返回一个实现了ISupervisor接口的实例。
(defn standalone-supervisor []
 (let [conf-atom (atom nil)
       id-atom (atom nil)]
   (reify ISupervisor
     ;; prepare方法主要功能是创建一个基于磁盘的存放K/V对的database——LocalState对象,LocalState类参见其定义部分
     (prepare [this conf local-dir]
       ;; conf-atom原子类型,绑定storm集群配置信息
       (reset! conf-atom conf)
   ;; state绑定LocalState对象,local-dir标识database在磁盘上的根目录,database实际就是一个HashMap对象序列化后存放到磁盘local-dir目录下
       (let [state (LocalState. local-dir)
         ;; LS-ID值为字符串"supervisor-id",定义在common.clj文件中。如果state中存放了该supervisor的id,那么curr-id绑定该id,否则curr-id绑定32为uuid
             curr-id (if-let [id (.get state LS-ID)]
                       id
           ;; 调用uuid函数生成一个32的id
                       (generate-supervisor-id))]
         ;; 调用state的put函数,更新该supervisor的id
         (.put state LS-ID curr-id)
     ;; id-atom原子类型,绑定该supervisor的id
         (reset! id-atom curr-id))
       )
     ;; 返回true
     (confirmAssigned [this port]
       true)
     ;; 从storm配置信息中获取supervisor的所有端口,因为clojure中的map函数返回的是"懒惰序列",所以需要调用doall函数对"懒惰序列"进行完全实例化
     (getMetadata [this]
       (doall (map int (get @conf-atom SUPERVISOR-SLOTS-PORTS))))
     ;; 获取supervisor的id
     (getSupervisorId [this]
       @id-atom)
     ;; 获取supervisor的分配id即其id
     (getAssignmentId [this]
       @id-atom)
     ;; killedWorker空实现
     (killedWorker [this port]
       )
     ;; assigned空实现
     (assigned [this ports]
       ))))

LocalState类是个java类,定义见LocalState.java,这个类有一个VersionedStore类型对象,VersionedStore类见VersionedStore.java,由于这两个类是java实现,而且也比较简单,这样就在详细分析。

mk-supervisor函数定义如下:

mk-supervisor函数

 

;; conf绑定storm集群配置信息,isupervisor绑定standalone-supervisor函数返回的实现ISupervisor接口的实例
(defserverfn mk-supervisor [conf shared-context ^ISupervisor isupervisor]
 ;; 打印日志信息
 (log-message "Starting Supervisor with conf " conf)
 ;; supervisor-isupervisor-dir函数调用了supervisor-local-dir函数,supervisor-local-dir函数从storm配置中获取storm的安装路径,然后在supervisor上创建目录{storm.local.dir}/supervisor,并返回目录
 ;; supervisor-isupervisor-dir函数返回字符串"{storm.local.dir}/supervisor/isupervisor"作为一个LocalState对象的根目录,该LocalState对象只用来存放supervisor的id
 ;; 调用isupervisor的prepare方法,创建一个LocalState对象,并生成该supervisor的id,将其存入LocalState对象
 (.prepare isupervisor conf (supervisor-isupervisor-dir conf))
 ;; supervisor-tmp-dir函数在supervisor上创建{storm.local.dir}/supervisor/tmp目录;FileUtils类的cleanDirectory方法清空该目录
 (FileUtils/cleanDirectory (File. (supervisor-tmp-dir conf)))
 ;; supervisor绑定supervisor元数据信息,supervisor-data参见其定义部分
 (let [supervisor (supervisor-data conf shared-context isupervisor)
   ;; event-manager和processes-event-manager分别绑定一个EventManager实例,managers绑定包含两个EventManager实例的集合。event-manager函数请参见文章"storm事件管理器EventManager源码分析-event.clj"
       [event-manager processes-event-manager :as managers] [(event/event-manager false) (event/event-manager false)]
   ;; partial用于定义"偏函数",所谓偏函数就是给一个指定函数的某些参数预赋值,这样就得到了一个新函数。sync-processes就绑定这个新函数,sync-processes参见其定义部分
       sync-processes (partial sync-processes supervisor)
   ;; synchronize-supervisor绑定一个函数,该函数主要功能就是当assignment发生变化时, 从nimbus同步topology的代码到本地,当assignment发生变化时, check workers状态, 保证被分配的work的状态都是valid
       synchronize-supervisor (mk-synchronize-supervisor supervisor sync-processes event-manager processes-event-manager)
   ;; heartbeat-fn绑定一个匿名函数,该匿名函数主要功能是调用StormClusterState实例的supervisor-heartbeat!函数将该supervisor的心跳信息SupervisorInfo实例写入zookeeper的"/supervisors/supervisor-id"节点中
       heartbeat-fn (fn [] (.supervisor-heartbeat!
                          ;; StormClusterState实例
                              (:storm-cluster-state supervisor)
                  ;; supervisor-id
                              (:supervisor-id supervisor)
                  ;; 创建SupervisorInfo实例,即该supervisor的心跳信息
                              (SupervisorInfo. (current-time-secs)
                                   ;; 主机名
                                               (:my-hostname supervisor)
                       ;; assignment-id即supervisor-id
                                               (:assignment-id supervisor)
                       ;; 当前集群已使用的所有port
                                               (keys @(:curr-assignment supervisor))
                                               ;; used ports
                       ;; 该supervisor上所有可用port,即在storm配置文件中配置的port
                                               (.getMetadata isupervisor)
                                               (conf SUPERVISOR-SCHEDULER-META)
                       ;; supervisor启动时间
                                               ((:uptime supervisor)))))]
   ;; 调用heartbeat-fn绑定的匿名函数,将supervisor心跳心跳写入zookeeper
   (heartbeat-fn)
   ;; should synchronize supervisor so it doesn‘t launch anything after being down (optimization)
   ;; 调用timer.clj中的schedule-recurring函数向该supervisor的定时器中添加一个周期执行的定时任务heartbeat-fn--"向zookeeper汇报superior的心跳信息",关于storm定时器的详细信息请参看"storm定时器timer源码分析-timer.clj"
   (schedule-recurring (:timer supervisor)
                       0
           ;; 指定每隔多长时间汇报一次心跳信息
                       (conf SUPERVISOR-HEARTBEAT-FREQUENCY-SECS)
                       heartbeat-fn)
   ;; 如果supervisor.enable值为true时(默认值就是true,而且不会改变,所以一定会执行),那么将synchronize-supervisor绑定的函数(mk-synchronize-supervisor函数返回的函数)每隔10s加入event-manager事件管理器中,
   ;; 这样即使zookeeper的"watcher机制"异常时,supervisor也可以主动的获取分配信息的变化。同时将sync-processes绑定的函数(sync-processes函数)每隔SUPERVISOR-MONITOR-FREQUENCY-SECS秒加入processes-event-manager事件管理器中,
   ;; 这样即使zookeeper的"watcher机制"异常时,supervisor也可以正常管理worker
   (when (conf SUPERVISOR-ENABLE)
     ;; This isn‘t strictly necessary, but it doesn‘t hurt and ensures that the machine stays up
     ;; to date even if callbacks don‘t all work exactly right
     (schedule-recurring (:timer supervisor) 0 10 (fn [] (.add event-manager synchronize-supervisor)))
     (schedule-recurring (:timer supervisor)
                         0
                         (conf SUPERVISOR-MONITOR-FREQUENCY-SECS)
                         (fn [] (.add processes-event-manager sync-processes))))
   (log-message "Starting supervisor with id " (:supervisor-id supervisor) " at host " (:my-hostname supervisor))
   ;; 返回实现了Shutdownable接口、SupervisorDaemon协议和DaemonCommon协议的实例
   (reify
    Shutdownable
    ;; 关闭supervisor,就是关闭该supervisor所拥护的资源
    (shutdown [this]
              (log-message "Shutting down supervisor " (:supervisor-id supervisor))
              (reset! (:active supervisor) false)
              (cancel-timer (:timer supervisor))
              (.shutdown event-manager)
              (.shutdown processes-event-manager)
              (.disconnect (:storm-cluster-state supervisor)))
    SupervisorDaemon
    ;; 返回集群配置信息
    (get-conf [this]
      conf)
    ;; 返回supervisor-id
    (get-id [this]
      (:supervisor-id supervisor))
    ;; 见名知意,关闭所有worker
    (shutdown-all-workers [this]
      (let [ids (my-worker-ids conf)]
        (doseq [id ids]
          (shutdown-worker supervisor id)
          )))
    DaemonCommon
    (waiting? [this]
      (or (not @(:active supervisor))
          (and
       ;; 定时器线程是否处于sleep状态
           (timer-waiting? (:timer supervisor))
       ;; 调用事件管理器的waiting?函数检查event-manager和processes-event-manager内事件执行线程是否处于sleep状态,memfn宏可以自动生成代码以使得java方法可以当成clojure里面的函数
           (every? (memfn waiting?) managers)))
          ))))

supervisor-data函数定义如下:

supervisor-data函数返回一个包含了supervisor元数据的map对象。

supervisor-data函数

 

(defn supervisor-data [conf shared-context ^ISupervisor isupervisor]
 ;; 保存集群配置信息
 {:conf conf
  ;; 启动supervisor时,shared-context为nil
  :shared-context shared-context
  ;; 保存supervisor实例
  :isupervisor isupervisor
  ;; 保存supervisor是否是活跃的(默认是活跃的)
  :active (atom true)
  ;; 保存supervisor启动时间
  :uptime (uptime-computer)
  ;; 保存工作线程id
  :worker-thread-pids-atom (atom {})
  ;; 保存StormClusterState对象
  :storm-cluster-state (cluster/mk-storm-cluster-state conf)
  ;; 保存supervisor的LocalState对象,该LocalState对象的根目录是"{storm.local.dir}/supervisor/localstate"
  :local-state (supervisor-state conf)
  ;; 保存supervisor的id
  :supervisor-id (.getSupervisorId isupervisor)
  ;; 保存supervisor的分配id,分配id与supervisor_id相同
  :assignment-id (.getAssignmentId isupervisor)
  ;; 保存supervisor的主机名,如果配置conf(map对象)中包含"storm.local.hostname",那么就使用配置的主机名,否则通过调用InetAddress.getLocalHost().getCanonicalHostName()获取主机名
  :my-hostname (if (contains? conf STORM-LOCAL-HOSTNAME)
                 (conf STORM-LOCAL-HOSTNAME)
                 (local-hostname))
  ;; 心跳时汇报当前集群的所有分配信息
  :curr-assignment (atom nil) ;; used for reporting used ports when heartbeating
  ;; 保存一个storm定时器timer,kill-fn函数会在timer-thread发生exception的时候被调用
  :timer (mk-timer :kill-fn (fn [t]
                              (log-error t "Error when processing event")
                              (exit-process! 20 "Error when processing an event")
                              ))
  ;; 创建一个用于存放带有版本号的分配信息的map
  :assignment-versions (atom {})
  })

sync-processes函数定义如下:

sync-processes函数

 

;; sync-processes函数用于管理workers, 比如处理不正常的worker或dead worker, 并创建新的workers
;; supervisor标识supervisor的元数据
(defn sync-processes [supervisor]
 ;; conf绑定storm的配置信息map
 (let [conf (:conf supervisor)
       ;; local-state绑定supervisor的LocalState实例
       ^LocalState local-state (:local-state supervisor)
       ;; 从supervisor的LocalState实例中获取本地分配信息端口port->LocalAssignment实例的map,LocalAssignment实例封装了storm-id和分配给该storm-id的executors
       assigned-executors (defaulted (.get local-state LS-LOCAL-ASSIGNMENTS) {})
       ;; now绑定当前时间
       now (current-time-secs)
       ;; allocated绑定worker-id->worker状态和心跳的map,read-allocated-workers函数请参见其定义部分
       allocated (read-allocated-workers supervisor assigned-executors now)
       ;; 过滤掉allocated中state不等于:valid的元素,并将过滤后的结果绑定到keepers
       keepers (filter-val
                (fn [[state _]] (= state :valid))
                allocated)
       ;; keep-ports绑定keepers中心跳信息所包含的端口
       keep-ports (set (for [[id [_ hb]] keepers] (:port hb)))
       ;; reassign-executors绑定assigned-executors中端口不在集合keep-ports的键值对构成的map,也就是说已分配的线程所对应的进程挂掉了,需要重新进行分配
       reassign-executors (select-keys-pred (complement keep-ports) assigned-executors)
       ;; new-worker-ids绑定port->worker-id的map,new-worker-ids保存了需要重新启动进程的worker-id
       new-worker-ids (into
                       {}
                       (for [port (keys reassign-executors)]
                         [port (uuid)]))
       ]
   ;; 1. to kill are those in allocated that are dead or disallowed
   ;; 2. kill the ones that should be dead
   ;;     - read pids, kill -9 and individually remove file
   ;;     - rmr heartbeat dir, rmdir pid dir, rmdir id dir (catch exception and log)
   ;; 3. of the rest, figure out what assignments aren‘t yet satisfied
   ;; 4. generate new worker ids, write new "approved workers" to LS
   ;; 5. create local dir for worker id
   ;; 5. launch new workers (give worker-id, port, and supervisor-id)
   ;; 6. wait for workers launch
 
   (log-debug "Syncing processes")
   (log-debug "Assigned executors: " assigned-executors)
   (log-debug "Allocated: " allocated)
   ;; allocated绑定worker-id->worker状态和心跳的map,id绑定worker-id,state绑定worker状态,heartbeat绑定worker心跳时间
   (doseq [[id [state heartbeat]] allocated]
     ;; 如果worker的状态不是:valid,那么就关闭worker
     (when (not= :valid state)
       (log-message
        "Shutting down and clearing state for id " id
        ". Current supervisor time: " now
        ". State: " state
        ", Heartbeat: " (pr-str heartbeat))
       ;; shutdown-worker函数关闭进程,shutdown-worker函数请参见其定义部分
       (shutdown-worker supervisor id)
       ))
   ;; new-worker-ids保存了需要重新启动进程的worker-id,遍历new-worker-ids,为每个worker-id创建本地目录"{storm.local.dir}/workers/{worker_id}"
   (doseq [id (vals new-worker-ids)]
     (local-mkdirs (worker-pids-root conf id)))
   ;; 将合并后的map重新保存到local-state的LS-APPROVED-WORKERS中
   (.put local-state LS-APPROVED-WORKERS
         ;; 将new-worker-ids的键值交换由原来的port->worker-id转换成worker-id->port,并与local-state的LS-APPROVED-WORKERS合并
         (merge
          ;; select-keys函数从local-state的LS-APPROVED-WORKERS中获取key包含在keepers中的键值对,返回结果是一个map
          (select-keys (.get local-state LS-APPROVED-WORKERS)
                       (keys keepers))
          ;; zipmap函数返回new-worker-ids的worker-id->port的map
          (zipmap (vals new-worker-ids) (keys new-worker-ids))
          ))
   ;; wait-for-workers-launch函数等待所有worker启动完成,请参见wait-for-workers-launch函数定义部分
   (wait-for-workers-launch
    conf
    ;; assignment绑定在该port运行的executor信息
    (dofor [[port assignment] reassign-executors]
      ;; id为port所对应的worker-id
      (let [id (new-worker-ids port)]
        (log-message "Launching worker with assignment "
                     (pr-str assignment)
                     " for this supervisor "
                     (:supervisor-id supervisor)
                     " on port "
                     port
                     " with id "
                     id
                     )
        ;; launch-worker函数负责启动worker,关于worker启动的相关分析会在以后的文章中详细介绍,在此不再介绍
        (launch-worker supervisor
                       (:storm-id assignment)
                       port
                       id)
        id)))
   ))

read-allocated-workers函数定义如下:

read-allocated-workers函数

 

;; 返回worker-id->worker状态和心跳的map,如果worker心跳为nil,那么worker是"dead"
(defn read-allocated-workers
 "Returns map from worker id to worker heartbeat. if the heartbeat is nil, then the worker is dead (timed out or never wrote heartbeat)"
 ;; supervisor绑定supervisor元数据,assigned-executors绑定supervisor分配信息端口port->LocalAssignment实例的map,now绑定当前时间
 [supervisor assigned-executors now]
 ;; 获取集群配置信息
 (let [conf (:conf supervisor)
   ;; 获取supervisor的LocalState实例
       ^LocalState local-state (:local-state supervisor)
   ;; id->heartbeat绑定supervisor上运行进程的worker-id->心跳信息的map
       id->heartbeat (read-worker-heartbeats conf)
   ;; approved-ids绑定supervisor的LocalState实例中保存的worker-id的集合
       approved-ids (set (keys (.get local-state LS-APPROVED-WORKERS)))]
   ;; 生成worker-id->[state hb]的map
   (into
    {}
    (dofor [[id hb] id->heartbeat]
           ;; cond相当于if...else嵌套
           (let [state (cond
                ;; 如果心跳信息为nil,那么state值为:not-started关键字
                        (not hb)
                          :not-started
            ;; 如果approved-ids不包含id或者matches-an-assignment?返回false,那么state值为:disallowed关键字
                        (or (not (contains? approved-ids id))
                ;; matches-an-assignment?函数通过比较心跳信息和分配信息中的storm-id和线程id集合是否相同,来判定该worker是否已分配
                            (not (matches-an-assignment? hb assigned-executors)))
                          :disallowed
            ;; 如果当前时间-上次心跳时间>心跳超时时间,state值为:timed-out关键字
                        (> (- now (:time-secs hb))
                           (conf SUPERVISOR-WORKER-TIMEOUT-SECS))
                          :timed-out
            ;; 以上条件均不满足时,state值为:valid关键字
                        true
                          :valid)]
             (log-debug "Worker " id " is " state ": " (pr-str hb) " at supervisor time-secs " now)
             [id [state hb]]
             ))
    )))

read-worker-heartbeats函数定义如下:

read-worker-heartbeats函数

 

;; 获取supervisor上运行进程的worker-id->心跳信息的map
(defn read-worker-heartbeats
 "Returns map from worker id to heartbeat"
 [conf]
 ;; ids绑定supervisor上进程的worker-id集合
 (let [ids (my-worker-ids conf)]
   ;; 生成worker-id->心跳信息的map
   (into {}
     (dofor [id ids]
       ;; read-worker-heartbeat函数获取指定worker-id的心跳信息,从supervisor上"{storm.local.dir}/workers/{worker-id}/heartbeats"中获取心跳信息
       [id (read-worker-heartbeat conf id)]))
   ))

my-worker-ids函数定义如下:

my-worker-ids函数

 

;; 获取supervisor上运行的进程的worker-id
(defn my-worker-ids [conf]
 ;; worker-root函数返回supervisor本地目录"{storm.local.dir}/workers",read-dir-contents函数获取目录"{storm.local.dir}/workers"下所有文件名的集合(即该supervisor上正在运行的所有进程的worker-id)
 (read-dir-contents (worker-root conf)))

matches-an-assignment?函数定义如下:

matches-an-assignment?函数

 

;; worker-heartbeat标识心跳信息,assigned-executors标识supervisor分配信息port->LocalAssignment实例的map
(defn matches-an-assignment? [worker-heartbeat assigned-executors]
 ;; 从worker-heartbeat中获取进程占用的端口,进而从assigned-executors中获取LocalAssignment实例
 (let [local-assignment (assigned-executors (:port worker-heartbeat))]
   ;; 如果local-assignment不为nil,且心跳信息中的storm-id和分配信息中的storm-id相等,且心跳信息中的线程id集合和分配信息中的线程id集合相等,那么返回true;否则返回false
   (and local-assignment
        (= (:storm-id worker-heartbeat) (:storm-id local-assignment))
    ;; Constants/SYSTEM_EXECUTOR_ID标识"系统bolt"的线程id,我定义的topology除了我们指定的spout和bolt外,还包含一些"系统bolt"
        (= (disj (set (:executors worker-heartbeat)) Constants/SYSTEM_EXECUTOR_ID)
           (set (:executors local-assignment))))))

shutdown-worker函数定义如下:

shutdown-worker函数

 

;; 关闭进程,supervisor标识supervisor元数据,id标识worker_id
(defn shutdown-worker [supervisor id]
 (log-message "Shutting down " (:supervisor-id supervisor) ":" id)
 ;; conf绑定集群配置信
 (let [conf (:conf supervisor)
       ;; 注意当storm集群"分布式模式"运行时,supervisor的"{storm.local.dir}/workers/{worker_id}/pids"路径中存放了worker实际对应的jvm进程id
       ;; 从supervisor的"{storm.local.dir}/workers/{worker_id}/pids"路径获取进程id,worker_id标识我们指定的进程id,pids目录存放了该worker实际对应的jvm进程的id
       pids (read-dir-contents (worker-pids-root conf id))
       ;; 注意当storm集群"本地模式"运行时,supervisor元数据中关键字:worker-thread-pids-atom所对应的map用于存放worker_id->线程id集合的键值对
       ;; 先从supervisor元数据中获取worker-id(我们人为分配给worker的id)->jvm进程id的map,thread-pid实际上绑定的是worker的jvm进程id
       thread-pid (@(:worker-thread-pids-atom supervisor) id)]
   ;; 当thread-pid不为空时,kill掉该进程
   (when thread-pid
     ;; 调用backtype.storm.process-simulator中的kill-process函数kill掉进程
     (psim/kill-process thread-pid))
   ;; 遍历pids集合,kill掉每个进程
   (doseq [pid pids]
     ;; 调用backtype.storm.util中的kill-process-with-sig-term函数,kill-process-with-sig-term函数又调用了send-signal-to-process函数,send-signal-to-process函数实现比较简单就是执行系统命令"kill -15 pid",kill掉进程
     ;; 注意在创建worker进程时为worker进程指定了关闭回调函数,当调用"kill -15 pid"关闭worker进程时会触发回调函数执行,回调函数是在worker.clj的mk-worker函数中添加的
     (kill-process-with-sig-term pid))
   ;; 如果pids不为空,sleep 1秒,等着"清理函数"--关闭回调函数执行完毕
   (if-not (empty? pids) (sleep-secs 1)) ;; allow 1 second for execution of cleanup threads on worker.
   ;; 通过调用"kill -15 pid"命令未能关闭的进程,将通过调用force-kill-process函数关闭,force-kill-process函数只是调用了"kill -9 pid"命令
   (doseq [pid pids]
     (force-kill-process pid)
     (try
       ;; 删除"{storm.local.dir}/workers/{worker_id}/pids"
       (rmpath (worker-pid-path conf id pid))
       (catch Exception e))) ;; on windows, the supervisor may still holds the lock on the worker directory
   ;; try-cleanup-worker函数清理本地目录,try-cleanup-worker函数参见其定义部分
   (try-cleanup-worker conf id))
 (log-message "Shut down " (:supervisor-id supervisor) ":" id))

try-cleanup-worker函数定义如下:

try-cleanup-worker函数

 

;; 清理本地目录
(defn try-cleanup-worker [conf id]
 (try
   ;; 删除"{storm.local.dir}/workers/{worker_id}/heartbeats"目录
   (rmr (worker-heartbeats-root conf id))
   ;; this avoids a race condition with worker or subprocess writing pid around same time
   ;; 删除"{storm.local.dir}/workers/{worker_id}/pids"目录
   (rmpath (worker-pids-root conf id))
   ;; 删除"{storm.local.dir}/workers/{worker_id}"目录
   (rmpath (worker-root conf id))
 (catch RuntimeException e
   (log-warn-error e "Failed to cleanup worker " id ". Will retry later")
   )
 (catch java.io.FileNotFoundException e (log-message (.getMessage e)))
 (catch java.io.IOException e (log-message (.getMessage e)))
   ))

wait-for-workers-launch函数定义如下:

wait-for-workers-launch函数

 

;; wait-for-workers-launch函数等待所有worker启动完成
(defn- wait-for-workers-launch [conf ids]
 (let [start-time (current-time-secs)]
   (doseq [id ids]
     ;; 调用wait-for-worker-launch函数
     (wait-for-worker-launch conf id start-time))
   ))

wait-for-worker-launch函数定义如下:

wait-for-worker-launch函数

 

;; wait-for-worker-launch函数等待worker启动完成,worker启动完成的条件是:如果worker在规定的心跳超时时间内有一次心跳那么就说明worker成功启动
(defn- wait-for-worker-launch [conf id start-time]
 (let [state (worker-state conf id)]    
   (loop []
     (let [hb (.get state LS-WORKER-HEARTBEAT)]
       (when (and
              (not hb)
              (<
               (- (current-time-secs) start-time)
               (conf SUPERVISOR-WORKER-START-TIMEOUT-SECS)
               ))
         (log-message id " still hasn‘t started")
         (Time/sleep 500)
         (recur)
         )))
   (when-not (.get state LS-WORKER-HEARTBEAT)
     (log-message "Worker " id " failed to start")
     )))

mk-synchronize-supervisor函数定义如下:

mk-synchronize-supervisor函数

 

;; mk-synchronize-supervisor函数返回一个名字为"this"的函数,
(defn mk-synchronize-supervisor [supervisor sync-processes event-manager processes-event-manager]
 (fn this []
   ;; conf绑定集群配置信息
   (let [conf (:conf supervisor)
         ;; storm-cluster-state绑定StormClusterState对象
         storm-cluster-state (:storm-cluster-state supervisor)
         ;; isupervisor绑定实现了ISupervisor接口的实例
         ^ISupervisor isupervisor (:isupervisor supervisor)
         ;; local-state绑定LocalState实例
         ^LocalState local-state (:local-state supervisor)
         ;; sync-callback绑定一个匿名函数,这个匿名函数的主要功能就是将上面定义的"this"函数添加到event-manager中,这样"this"函数将会在一个新的线程内执行
         ;; 每次执行,都需要再一次把sync-callback注册到zookeeper中作为回调函数,以保证下次可以被继续触发,当zookeeper的子节点"/assignments"发生变化时执行回调函数sync-callback
         sync-callback (fn [& ignored] (.add event-manager this))
         ;; assignment-versions绑定带有版本号的分配信息,topology-id->分配信息的map
         assignment-versions @(:assignment-versions supervisor)
         ;; assignments-snapshot绑定topoloy-id->分配信息AssignmentInfo对象的map,versions绑定带有版本号的分配信息,assignments-snapshot函数从zookeeper的子节点"/assignments"获取分配信息(当前集群分配信息快照),并将回调函数添加到子节点"/assignments"上,assignments-snapshot函数参见其定义部分
         {assignments-snapshot :assignments versions :versions}  (assignments-snapshot
                                                                  storm-cluster-state sync-callback
                                                                  assignment-versions)
         ;; 调用read-storm-code-locations函数获取topology-id->nimbus上该topology代码目录的map                                                
         storm-code-map (read-storm-code-locations assignments-snapshot)
         ;; read-downloaded-storm-ids函数从supervisor本地的"{storm.local.dir}/stormdist"目录读取已经下载了代码jar包的topology-id
         downloaded-storm-ids (set (read-downloaded-storm-ids conf))
         ;; all-assignment绑定该supervisor上的所有分配信息,即port->LocalAssignment对象的map
         all-assignment (read-assignments
                          assignments-snapshot
                          (:assignment-id supervisor))
         ;; 调用isupervisor对象的confirmAssigned函数验证all-assignment的key即port的有效性,将通过验证的保存到new-assignment中。isupervisor对象是在standalone-supervisor函数中创建的,查看standalone-supervisor函数,我们可以发现isupervisor对象的confirmAssigned函数只是返回true,所以new-assignment=all-assignment
         new-assignment (->> all-assignment
                             (filter-key #(.confirmAssigned isupervisor %)))
         ;; assigned-storm-ids绑定分配给该supervisor的topology-id的集合
         assigned-storm-ids (assigned-storm-ids-from-port-assignments new-assignment)
         ;; existing-assignment绑定该supervisor上已经存在的分配信息
         existing-assignment (.get local-state LS-LOCAL-ASSIGNMENTS)]
     (log-debug "Synchronizing supervisor")
     (log-debug "Storm code map: " storm-code-map)
     (log-debug "Downloaded storm ids: " downloaded-storm-ids)
     (log-debug "All assignment: " all-assignment)
     (log-debug "New assignment: " new-assignment)
     
     ;; download code first
     ;; This might take awhile
     ;;   - should this be done separately from usual monitoring?
     ;; should we only download when topology is assigned to this supervisor?
     ;; storm-code-map绑定当前集群上已分配的所有topology-id->nimbus上代码jar包目录的键值对的map
     (doseq [[storm-id master-code-dir] storm-code-map]
       ;; 如果downloaded-storm-ids集合不包含该storm-id,且assigned-storm-ids集合包含该storm-id(表明该storm-id需要在该superior上运行,但是该storm-id的代码jar包还没有从nimbus服务器下载到本地),则调用download-storm-code函数下载代码jar包
       (when (and (not (downloaded-storm-ids storm-id))
                  (assigned-storm-ids storm-id))
         (log-message "Downloading code for storm id "
            storm-id
            " from "
            master-code-dir)
         ;; 从nimbus服务器上下载该storm-id相关的代码jar包,序列化后的topology对象,运行时所需的配置信息,并将其保存到"{storm.local.dir}/nimbus/stormdist/{storm-id}/"目录
         (download-storm-code conf storm-id master-code-dir)
         (log-message "Finished downloading code for storm id "
            storm-id
            " from "
            master-code-dir)
         ))

     (log-debug "Writing new assignment "
                (pr-str new-assignment))
     ;; existing-assignment与new-assignment的差集表示不需要在该supervisor上运行的分配的集合,所以要把这些分配对应的worker关闭
     (doseq [p (set/difference (set (keys existing-assignment))
                               (set (keys new-assignment)))]
       ;; 当前storm版本0.9.2中,killedWorker为空实现,所以什么都没做
       (.killedWorker isupervisor (int p)))
     ;; assigned函数为空实现,什么也没有做
     (.assigned isupervisor (keys new-assignment))
     ;; 将最新分配信息new-assignment保存到local-state数据库中
     (.put local-state
           LS-LOCAL-ASSIGNMENTS
           new-assignment)
     ;; 将带有版本号的分配信息versions存入supervisor缓存:assignment-versions中
     (swap! (:assignment-versions supervisor) versions)
     ;; 重新设置supervisor缓存的:curr-assignment值为new-assignment,即保存当前storm集群上最新分配信息
     (reset! (:curr-assignment supervisor) new-assignment)
     ;; remove any downloaded code that‘s no longer assigned or active
     ;; important that this happens after setting the local assignment so that
     ;; synchronize-supervisor doesn‘t try to launch workers for which the
     ;; resources don‘t exist
     ;; 如果当前supervisor服务器的操作系统是"Windows_NT"系统,那么执行shutdown-disallowed-workers函数,关闭状态为:disallowed的worker
     (if on-windows? (shutdown-disallowed-workers supervisor))
     ;; 遍历downloaded-storm-ids集合,该集合内存放了已经下载了jar包等信息的topology的id
     (doseq [storm-id downloaded-storm-ids]
       ;; 如果storm-id不在assigned-storm-ids集合内,则递归删除"{storm.local.dir}/supervisor/stormdist/{storm-id}"目录。assigned-storm-ids表示当前需要在该supervisor上运行的topology的id
       (when-not (assigned-storm-ids storm-id)
         (log-message "Removing code for storm id "
                      storm-id)
         (try
           (rmr (supervisor-stormdist-root conf storm-id))
           (catch Exception e (log-message (.getMessage e))))
         ))
     ;; 将sync-processes函数添加到processes-event-manager事件管理器中,这样就可以在一个单独线程内执行sync-processes函数。因为sync-processes函数比较耗时,所以需要在一个新的线程内执行
     (.add processes-event-manager sync-processes)
     )))

assignments-snapshot函数定义如下:

assignments-snapshot函数

 

;; assignments-snapshot函数从zookeeper的子节点"/assignments"获取分配信息,并将回调函数添加到子节点"/assignments"上,assignment-versions绑定该supervisor本地缓存的带有版本号的分配信息
(defn- assignments-snapshot [storm-cluster-state callback assignment-versions]
 ;; storm-ids绑定已分配的topology-id的集合,获取/assignments的子节点列表,如果callback不为空,将其赋值给assignments-callback,并对/assignments添加"节点观察",这样supervisor就能感知集群是否有新的assignment或者有assignment被删除
 (let [storm-ids (.assignments storm-cluster-state callback)]
   ;; new-assignments绑定最新分配信息
   (let [new-assignments
         (->>
          ;; sid绑定topology-id
          (dofor [sid storm-ids]
                 ;; recorded-version绑定该supervisor上缓存的该sid的分配信息版本号
                 (let [recorded-version (:version (get assignment-versions sid))]
                   ;; assignment-version绑定zookeeper上"/assignments/{sid}"节点数据及其版本号,并注册回调函数
                   (if-let [assignment-version (.assignment-version storm-cluster-state sid callback)]
                     ;; 如果缓存的分配版本号和zookeeper上获取的分配版本号相等,则返回sid->缓存的分配信息的map,否则从zookeeper的"/assignments/{sid}"节点重新获取带有版本号的分配信息,并注册回调函数,这样supervisor就能感知某个已存在的assignment是否被重新分配
                     (if (= assignment-version recorded-version)
                       {sid (get assignment-versions sid)}
                       {sid (.assignment-info-with-version storm-cluster-state sid callback)})
                     ;; 如果从zookeeper上获取分配信息失败,值为{sid nil}
                     {sid nil})))
          ;; 将dofor结果进行合并,形如:{sid_1 {:data data_1 :version version_1}, sid_2 {:data data_2 :version version_2},......sid_n {:data data_n :version version_n} }
          (apply merge)
          ;; 保留值不空的键值对
          (filter-val not-nil?))]
     ;; 返回的map形如:{:assignments {sid_1 data_1, sid_2 data_2, ...... , sid_n data_n}, :versions {sid_1 {:data data_1 :version version_1}, sid_2 {:data data_2 :version version_2},......sid_n {:data data_n :version version_n} } }    
     ;; data_x是一个AssignmentInfo对象,AssignmentInfo对象包含对应的nimbus上的代码目录,所有task的启动时间,每个task与机器、端口的映射
     {:assignments (into {} (for [[k v] new-assignments] [k (:data v)]))
      :versions new-assignments})))

read-assignments函数定义如下:

read-assignments函数

 

(defn- read-assignments
 "Returns map from port to struct containing :storm-id and :executors"
 ;; assignments-snapshot绑定topology-id->分配信息AssignmentInfo对象的map,assignment-id绑定supervisor-id
 [assignments-snapshot assignment-id]
 ;; 遍历read-my-executors函数返回结果,检查是否存在多个topology分配到同一个端口,如果存在则抛出异常。检查的方式特别巧妙,通过对返回结果调用merge-with函数,如果返回结果中存在相同的port,那么就会调用
 ;; 匿名函数(fn [& ignored] ......),这样就会抛出异常
 (->> (dofor [sid (keys assignments-snapshot)] (read-my-executors assignments-snapshot sid assignment-id))
      (apply merge-with (fn [& ignored] (throw-runtime "Should not have multiple topologies assigned to one port")))))

read-my-executor函数定义如下:

read-my-executor函数

 

;; assignments-snapshot绑定topology-id->分配信息AssignmentInfo对象的map,assignment-id绑定supervisor-id,storm-id为topoloy-id
(defn- read-my-executors [assignments-snapshot storm-id assignment-id]
 (let [assignment (get assignments-snapshot storm-id)
       ;; my-executors绑定分配给该supervisor的executor信息,即executor->node+port的map
       my-executors (filter (fn [[_ [node _]]] (= node assignment-id))
                          (:executor->node+port assignment))
       ;; port-executors绑定port->executor-id集合的map,merge-with函数的作用就是对key相同的value调用concat函数
       port-executors (apply merge-with
                         concat
                         (for [[executor [_ port]] my-executors]
                           {port [executor]}
                           ))]
   ;; 返回port->LocalAssignment对象的map,LocalAssignment包含两个属性:topology-id和executor-id集合
   (into {} (for [[port executors] port-executors]
              ;; need to cast to int b/c it might be a long (due to how yaml parses things)
              ;; doall is to avoid serialization/deserialization problems with lazy seqs
              [(Integer. port) (LocalAssignment. storm-id (doall executors))]
              ))))

download-storm-code函数定义如下:

download-storm-code函数

 

(defmulti download-storm-code cluster-mode)
download-storm-code函数是一个"多重函数",根据cluster-mode函数的返回值决定调用哪个函数,cluster-mode函数可能返回关键字:distributed:local,如果返回:distributed,那么会调用下面这个函数。
(defmethod download-storm-code
   ;; master-code-dir绑定storm-id的代码jar包在nimbus服务器上的路径
   :distributed [conf storm-id master-code-dir]
   ;; Downloading to permanent location is atomic
   ;; tmproot绑定supervisor本地路径"{storm.local.dir}/supervisor/tmp/{uuid}",临时存放从nimbus上下载的代码jar包
   (let [tmproot (str (supervisor-tmp-dir conf) file-path-separator (uuid))
         ;; stormroot绑定该storm-id的代码jar包在supervisor上的路径"{storm.local.dir}/supervisor/stormdist/{storm-id}"
         stormroot (supervisor-stormdist-root conf storm-id)]
     ;; 创建临时目录tmproot
     (FileUtils/forceMkdir (File. tmproot))
     ;; 将nimbus服务器上的"{storm.local.dir}/nimbus/stormdist/{storm-id}/stormjar.jar"文件下载到supervisor服务器的tmproot目录中,stormjar.jar包含这个topology所有代码
     (Utils/downloadFromMaster conf (master-stormjar-path master-code-dir) (supervisor-stormjar-path tmproot))
     ;; 将nimbus服务器上的"{storm.local.dir}/nimbus/stormdist/{storm-id}/stormcode.ser"文件下载到supervisor服务器的tmproot目录中,stormcode.ser是这个topology对象的序列化
     (Utils/downloadFromMaster conf (master-stormcode-path master-code-dir) (supervisor-stormcode-path tmproot))
     ;; 将nimbus服务器上的"{storm.local.dir}/nimbus/stormdist/{storm-id}/stormconf.ser"文件下载到supervisor服务器的tmproot目录中,stormconf.ser包含运行这个topology的配置
     (Utils/downloadFromMaster conf (master-stormconf-path master-code-dir) (supervisor-stormconf-path tmproot))
     ;; RESOURCES-SUBDIR值为字符串"resources",extract-dir-from-jar函数主要作用就是将jar包解压,然后将jar包中路径以"resources"开头的文件解压到"{tmproot}/resources/......"目录
     (extract-dir-from-jar (supervisor-stormjar-path tmproot) RESOURCES-SUBDIR tmproot)
     ;; 将临时目录tmproot中的文件剪切到stormroot目录中,这样"{storm.local.dir}/nimbus/stormdist/{storm-id}/"目录中将包括resources目录,stormjar.jar文件,stormcode.ser文件,stormconf.ser文件
     (FileUtils/moveDirectory (File. tmproot) (File. stormroot))
     ))

extract-dir-from-jar函数定义如下:

extract-dir-from-jar函数

 

;; jarpath标识jar路径,dir标识"resources",destdir标识"{tmproot}"路径
(defn extract-dir-from-jar [jarpath dir destdir]
 (try-cause
   ;; 使用类ZipFile来解压jar包,jarpath绑定ZipFile对象
   (with-open [jarpath (ZipFile. jarpath)]
     ;; 调用entries方法,返回一个枚举对象,然后调用enumeration-seq函数获取文件的ZIP条目对象
     (let [entries (enumeration-seq (.entries jarpath))]
       ;; 遍历entries中路径以"resources"开头的文件
       (doseq [file (filter (fn [entry](and (not (.isDirectory entry)) (.startsWith (.getName entry) dir))) entries)]
         ;; 在"tmproot"目录中创建文件的完整父路径
         (.mkdirs (.getParentFile (File. destdir (.getName file))))
         ;; 将文件复制到"{tmproot}/{在压缩文件中的路径}"
         (with-open [out (FileOutputStream. (File. destdir (.getName file)))]
           (io/copy (.getInputStream jarpath file) out)))))
   (catch IOException e
     (log-message "Could not extract " dir " from " jarpath))))

 

以上就是storm启动supervisor的完整流程,启动supervisor的工作主要是在mk-supervisor函数中进行的,所以阅读该部分源码时,要首先从该函数入手,然后依次分析在该函数中所调用的其他函数,根据函数的控制流程分析每个函数。

storm启动supervisor源码分析-supervisor.clj

标签:des   style   http   color   io   os   使用   java   ar   

原文地址:http://www.cnblogs.com/ierbar0604/p/3991488.html

(0)
(0)
   
举报
评论 一句话评论(0
登录后才能评论!
© 2014 mamicode.com 版权所有  联系我们:gaon5@hotmail.com
迷上了代码!