Question

用例：在Django框架（版本1.6）中使用celery来安排基本上写入数据库的任务。我只有一个自定义队列，芹菜节拍调度程序将任务放在上面。创建了一个监听此队列的芹菜工作者，该队列的并发性为8

问题：8名个体工作者中的每一个都开始创建永不回收的线程（我的猜测）。这导致太多线程（我已经看到计数达到20k线程）。在4-5小时的时间内，线程数接触10k！

错误我看到：无法启动新线程。

关于谁正在开始新线程的Python回溯给了我：调用django save创建了一个新线程。 “adgroup”这里是一个django模型对象

[2015-12-03 18:40:17,133: WARNING/Worker-3] adgroup.save(update_fields=['bids_today', 'impressions_today', 'spent_today', 'last_metric_update_time'])
[2015-12-03 18:40:17,887: WARNING/Worker-3] File "/home/ec2-user/venv/local/lib/python2.7/dist-packages/django/db/models/base.py", line 545, in save
[2015-12-03 18:40:17,887: WARNING/Worker-3] force_update=force_update, update_fields=update_fields)
[2015-12-03 18:40:18,715: WARNING/Worker-3] File "/home/ec2-user/venv/local/lib/python2.7/dist-packages/django/db/models/base.py", line 582, in save_base
[2015-12-03 18:40:18,716: WARNING/Worker-3] update_fields=update_fields, raw=raw, using=using)
[2015-12-03 18:40:18,716: WARNING/Worker-3] File "/home/ec2-user/venv/local/lib/python2.7/dist-packages/django/dispatch/dispatcher.py", line 185, in send
[2015-12-03 18:40:18,716: WARNING/Worker-3] response = receiver(signal=self, sender=sender, **named)
[2015-12-03 18:40:19,300: INFO/MainProcess] Task ExtendTV.celery_tasks.stats_collector.collectAdGroupMetricsTask[2ae52b3d-77b9-46d3-93ac-d7fad9b96382] succeeded in 26.486441362s: None
[2015-12-03 18:40:19,395: WARNING/Worker-3] File "/home/ec2-user/venv/local/lib/python2.7/dist-packages/haystack/signals.py", line 48, in handle_save
[2015-12-03 18:40:19,593: WARNING/Worker-3] index.update_object(instance, using=using)
[2015-12-03 18:40:19,593: WARNING/Worker-3] File "/home/ec2-user/venv/local/lib/python2.7/dist-packages/haystack/indexes.py", line 274, in update_object
[2015-12-03 18:40:19,593: WARNING/Worker-3] backend.update(self, [instance])
[2015-12-03 18:40:19,593: WARNING/Worker-3] File "/home/ec2-user/venv/local/lib/python2.7/dist-packages/haystack/backends/whoosh_backend.py", line 208, in update
[2015-12-03 18:40:20,515: WARNING/Worker-3] writer.commit()
[2015-12-03 18:40:20,516: WARNING/Worker-3] File "/home/ec2-user/venv/local/lib/python2.7/dist-packages/whoosh/writing.py", line 1043, in commit
[2015-12-03 18:40:21,318: WARNING/Worker-3] self.start()
[2015-12-03 18:40:21,642: WARNING/Worker-3] File "/usr/lib64/python2.7/threading.py", line 748, in start
[2015-12-03 18:40:22,340: WARNING/Worker-3] _start_new_thread(self.__bootstrap, ())
[2015-12-03 18:40:22,340: WARNING/Worker-3] error: can't start new thread

其他信息：从图中可以看出，内存完全在正常范围内。先前版本的celery 3.0.x中不存在此“线程问题”。然而，这里的记忆变得非常高

Celery命令我用来创建一个worker：

celery -A ProjectName worker -l DEBUG -Q ExampleQueueName

我使用的芹菜设置：

CELERY_DEFAULT_QUEUE = 'default'
CELERY_DEFAULT_EXCHANGE_TYPE = 'direct'
CELERY_DEFAULT_ROUTING_KEY = 'default'
CELERY_RESULT_SERIALIZER = 'json'
CELERY_ACCEPT_CONTENT = ['json']
CELERY_TASK_SERIALIZER = 'json'
CELERY_RESULT_SERIALIZER = 'json'
CELERY_TASK_RESULT_EXPIRES=60*60*24
CELERYD_PREFETCH_MULTIPLIER = 128

其他相关设置：使用rabbitmq 3.5.4作为消息代理

更新：

def collectAdGroupMetricsTask(*args, **kwargs):
    try:
        adgroup = AdGroup.objects.get(id=kwargs.get("adgroupID"))
        collectAdGroupMetrics(adgroup)
    except Exception as e:
        logger.error("Could not retreive AdGroup for collectAdGroupMetrics. " + str(e))
    return

def collectAdGroupMetrics(adgroup, currDate=None):
    Value1=function1_making_another_db_call()
    Value2=function2_making_another_db_call()
    adgroup.fieldname1 = Value1
    adgroup.fieldname2 = Value2    
    adgroup.save(update_fields=['fieldname1', 'fieldname2'])

Example of worker process having lots of threads.

Answer 1

whoosh（python包）试图获取写锁定并等待导致创建如此多的threads.Hence从django中已安装的应用程序列表中删除了whoosh。还在芹菜中使用maxtasksperchild配置来防止内存不断增长。

Answer 2

首先在pythonic虚拟环境中安装 gevent 软件包。
下一步对运行 celery
最后，我附加了参数--pool gevent。默认情况下， celery 使用池“ prefork”，该池应该有一些错误。
选择 celery 的进程数量后，降低到子进程（并发）的数量。

Celery 3.1.19为每个worker创建了太多线程，导致服务器超载100％CPU无法创建更多线程

2 个答案: