对于ambari-collector 部分源码流程的简单理解

工作中遇到了ambari指标的搜集和展示问题，因需要增添部分python脚本对平台的脚本数据进行汇聚和整理，所以需要理解ambari-collector的部分流程，所以进行了简要的阅读，故做此分析，以防止后续遗忘。以下代码都存在删减，仅供参考。

首先来看一下main.py

from core.controller import Controller
def main(argv=None):
  # Allow Ctrl-C
  stop_handler = bind_signal_handlers()
  server_process_main(stop_handler)
def server_process_main(stop_handler, scmStatus=None):
  if scmStatus is not None:
    scmStatus.reportStartPending()
  config = Configuration()
  _init_logging(config)
  controller = Controller(config, stop_handler)
  logger.info('Starting Server RPC Thread: %s' % ' '.join(sys.argv))
  controller.start()
  print "Server out at: " + SERVER_OUT_FILE
  print "Server log at: " + SERVER_LOG_FILE
  save_pid(os.getpid(), PID_OUT_FILE)
  if scmStatus is not None:
    scmStatus.reportStarted()
  #The controller thread finishes when the stop event is signaled
  controller.join()
  remove_file(PID_OUT_FILE)
  pass

由上述代码可以看出来，server_process_main 里面实例化了Controller类，并controller.start()开启了一个线程，那么我们看一下Controller里面的的代码（有删略）

from emitter import Emitter
class Controller(threading.Thread):
  def __init__(self, config, stop_handler):
    # Process initialization code
    threading.Thread.__init__(self)
    self.emitter = Emitter(self.config, self.application_metric_map, stop_handler)
  def run(self):
    self.start_emitter()
    Timer(1, self.addsecond).start()
    while True:
      if (self.event_queue.full()):
        logger.warn('Event Queue full!! Suspending further collections.')
      else:
        self.enqueque_events()
      pass
      if 0 == self._stop_handler.wait(self.sleep_interval):
        logger.info('Shutting down Controller thread')
        break
      if not self._t is None:
        self._t.cancel()
        self._t.join(5)
     self.emitter.join(5)
     pass
  def start_emitter(self):
    self.emitter.start()

run函数里面执行了start_emitter()，然后对Emitter进行实例化，执行了emmitter.start()，接下来我们看一下Emitter的代码

class Emitter(threading.Thread):
  COLLECTOR_URL = "xxxxx"
  RETRY_SLEEP_INTERVAL = 5
  MAX_RETRY_COUNT = 3
  def __init__(self, config, application_metric_map, stop_handler):
    threading.Thread.__init__(self)
    self.lock = threading.Lock()
    self.collector_address = config.get_server_address()
    self.send_interval = config.get_send_interval()
    self._stop_handler = stop_handler
    self.application_metric_map = application_metric_map
  def run(self):
    while True:
      try:
        self.submit_metrics()
      except Exception, e:
        logger.warn('Unable to emit events. %s' % str(e))
      pass
      if 0 == self._stop_handler.wait(self.send_interval):
        logger.info('Shutting down Emitter thread')
        return
    pass
def submit_metrics(self):
    retry_count = 0
    # This call will acquire lock on the map and clear contents before returning
    # After configured number of retries the data will not be sent to the
    # collector
    json_data = self.application_metric_map.flatten(None, True)
    if json_data is None:
      logger.info("Nothing to emit, resume waiting.")
      return
    pass
    response = None
    while retry_count < self.MAX_RETRY_COUNT:
      try:
        response = self.push_metrics(json_data)
      except Exception, e:
        logger.warn('Error sending metrics to server. %s' % str(e))
      pass
      if response and response.getcode() == 200:
        retry_count = self.MAX_RETRY_COUNT
      else:
        logger.warn("Retrying after {0} ...".format(self.RETRY_SLEEP_INTERVAL))
        retry_count += 1
        #Wait for the service stop event instead of sleeping blindly
        if 0 == self._stop_handler.wait(self.RETRY_SLEEP_INTERVAL):
          return
      pass
    pass
  def push_metrics(self, data):
    headers = {"Content-Type" : "application/json", "Accept" : "*/*"}
    server = self.COLLECTOR_URL.format(self.collector_address.strip())
    logger.info("server: %s" % server)
    logger.debug("message to sent: %s" % data)
    req = urllib2.Request(server, data, headers)
    response = urllib2.urlopen(req, timeout=int(self.send_interval - 10))
    if response:
      logger.debug("POST response from server: retcode = {0}".format(response.getcode()))
      logger.debug(str(response.read()))
    pass
    return response

由上述代码可以看出来，run函数执行的时候，执行了submit_metrics()函数，重点来了，该函数的核心就是 json_data = self.application_metric_map.flatten(None, True),当前类继承自ApplicationMetricsMap，让我们去查看一下ApplicationMetricsMap的代码

def flatten(self, application_id = None, clear_once_flattened = False):
    with self.lock:
      timeline_metrics = { "metrics" : [] }
      local_metric_map = {}
      if application_id:
        if self.app_metric_map.has_key(application_id):
          local_metric_map = { application_id : self.app_metric_map[application_id] }
        else:
          logger.info("application_id: {0}, not present in the map.".format(application_id))
      else:
        local_metric_map = self.app_metric_map.copy()
      pass
      for appId, metrics in local_metric_map.iteritems():
        for metricId, metricData in dict(metrics).iteritems():
          # Create a timeline metric object
          timeline_metric = {
            "hostname" : self.hostname if appId == "HOST" else "",
            "metricname" : metricId,
            #"appid" : "HOST",
            "appid" : appId,
            "instanceid" : "",
            "starttime" : self.get_start_time(appId, metricId),
            "metrics" : metricData
          }
          timeline_metrics[ "metrics" ].append( timeline_metric )
        pass
      pass
      json_data = json.dumps(timeline_metrics) if len(timeline_metrics[ "metrics" ]) > 0 else None
      if clear_once_flattened:
        self.app_metric_map.clear()
      pass
      return json_data
  pass

由此函数可以看得出来，该函数主要就是对数据进行一些合并，汇聚形成新的数据结构，但是当第一次在Controller里面执行start_emmiter()时候，该函数并未执行，因为self.app_metric_map的数据结构并未生成，让我们往前看，在Controller的run函数里面有这么一行代码，self.enqueue_events()，从字面意思看出来是事件入队列，让我们找到该函数，最终进行相互调用后是执行了process_service_collection_event

 def process_service_collection_event(self, event):
    startTime = int(round(time() * 1000))
    metrics = None
    path = os.path.abspath('.')
    for root, dirs, files in os.walk("%s/libs/" % path):
      appid = event.get_group_name().split('_')[0]
      metricgroup = event.get_group_name().split('_')[1]
      if ("%s_metrics.sh" % appid) in filter(lambda x: ".sh" in x, files):
        metrics = {appid: self.service_info.get_service_metrics(appid, metricgroup)}
      else:
        logger.warn('have no %s modules' % appid)
    if metrics:
      for item in metrics:
        self.application_metric_map.put_metric(item, metrics[item], startTime)
    pass

这段代码就是执行各个服务的脚本，然后汇聚数据，最终生成metrics变量，然后执行了self.application_metric_map.put_metric(item, metrics[item], startTime)，这个application_metric_map其实就是ApplicationMetricMap类的实例，其中有一个函数如下所示：

 def put_metric(self, application_id, metric_id_to_value_map, timestamp):
    with self.lock:
      for metric_name, value in metric_id_to_value_map.iteritems():
        metric_map = self.app_metric_map.get(application_id)
        if not metric_map:
          metric_map = { metric_name : { timestamp : value } }
          self.app_metric_map[ application_id ] = metric_map
        else:
          metric_id_map = metric_map.get(metric_name)
          if not metric_id_map:
            metric_id_map = { timestamp : value }
            metric_map[ metric_name ] = metric_id_map
          else:
            metric_map[ metric_name ].update( { timestamp : value } )
          pass
        pass
  pass

其实这段代码主要是从脚本中搜集的数据，形成最终的app_metric_map数据，在Controller中一直被无限调用，只是我们第一次执行start_emitter（）时候并未执行而已，当从脚本中搜集到数据，才会执行真正的调用，然后通过requests模块，上报到 metrics collector的6188端口中，最终数据落于hbase中。