[源码解析] 并行分布式框架 Celery 之 worker 启动 (2)
目录
0x00 摘要
Celery是一个简单、灵活且可靠的，处理大量消息的分布式系统，专注于实时处理的异步任务队列，同时也支持任务调度。Celery 是调用其Worker 组件来完成具体任务处理。
前文讲了 Celery 启动过程的前半部分，本文继续后半部分的分析。
本文相关其他链接如下：
以及
[源码解析] 并行分布式框架 Celery 之 worker 启动 (1)
0x01 前文回顾
前文提到了，我们经过一系列过程，正式来到了 Worker 的逻辑。我们在本文将接下来继续看后续 work as a program 的启动过程。
+----------------------+
+----------+ | @cached_property |
| User | | Worker |
+----+-----+ +---> | |
| | | |
| worker_main | | Worker application |
| | | celery/app/base.py |
v | +----------------------+
+---------+------------+ |
| Celery | |
| | |
| Celery application | |
| celery/app/base.py | |
| | |
+---------+------------+ |
| |
| celery.main |
| |
v |
+---------+------------+ |
| @click.pass_context | |
| celery | |
| | |
| | |
| CeleryCommand | |
| celery/bin/celery.py | |
| | |
+---------+------------+ |
| |
| |
| |
v |
+----------+------------+ |
| @click.pass_context | |
| worker | |
| | |
| | |
| WorkerCommand | |
| celery/bin/worker.py | |
+-----------+-----------+ |
| |
+-----------------+
为了便于大家理解，我们先给出最终的流程图如下：
0x2 Worker as a program
这里的 worker 其实就是 业务主体，值得大书特书。
代码来到了celery/apps/worker.py。
class Worker(WorkController):
"""Worker as a program."""
实例化的过程调用到了WorkController基类的init。
初始化基本就是：
- loader 加载各种配置；
- setup_defaults做缺省设置；
- setup_instance 就是正式建立，包括配置存放消息的queue。
- 通过Blueprint来建立 Worker 内部的各个子模块。
代码位于celery/apps/worker.py。
class WorkController:
"""Unmanaged worker instance."""
app = None
pidlock = None
blueprint = None
pool = None
semaphore = None
#: contains the exit code if a :exc:`SystemExit` event is handled.
exitcode = None
class Blueprint(bootsteps.Blueprint):
"""Worker bootstep blueprint."""
name = 'Worker'
default_steps = {
'celery.worker.components:Hub',
'celery.worker.components:Pool',
'celery.worker.components:Beat',
'celery.worker.components:Timer',
'celery.worker.components:StateDB',
'celery.worker.components:Consumer',
'celery.worker.autoscale:WorkerComponent',
}
def __init__(self, app=None, hostname=None, **kwargs):
self.app = app or self.app # 设置app属性
self.hostname = default_nodename(hostname) # 生成node的hostname
self.startup_time = datetime.utcnow()
self.app.loader.init_worker() # 调用app.loader的init_worker()方法
self.on_before_init(**kwargs) # 调用该初始化方法
self.setup_defaults(**kwargs) # 设置默认值
self.on_after_init(**kwargs)
self.setup_instance(**self.prepare_args(**kwargs)) # 建立实例
此时会调用app.loader的init_worker方法，
2.1 loader
此处的app.loader，是在Celery初始化的时候设置的loader属性，该值默认是celery.loaders.app:AppLoader。其作用就是导入各种配置。
其位于celery/loaders/base.py，定义如下：
@cached_property
def loader(self):
"""Current loader instance."""
return get_loader_cls(self.loader_cls)(app=self)
get_loader_cls如下：
def get_loader_cls(loader):
"""Get loader class by name/alias."""
return symbol_by_name(loader, LOADER_ALIASES, imp=import_from_cwd)
此时的loader实例就是AppLoader，然后调用该类的init_worker方法，
def init_worker(self):
if not self.worker_initialized: # 如果该类没有被设置过
self.worker_initialized = True # 设置成设置过
self.import_default_modules() # 导入默认的modules
self.on_worker_init()
import_default_modules如下，主要就是导入在app配置文件中需要导入的modules，
def import_default_modules(self):
responses = signals.import_modules.send(sender=self.app)
# Prior to this point loggers are not yet set up properly, need to
# check responses manually and reraised exceptions if any, otherwise
# they'll be silenced, making it incredibly difficult to debug.
for _, response in responses: # 导入项目中需要导入的modules
if isinstance(response, Exception):
raise response
return [self.import_task_module(m) for m in self.default_modules]
2.2 setup_defaults in worker
继续分析Worker类初始化过程中的self.setup_defaults方法，给运行中需要设置的参数设置值，
这之后，self.pool_cls的数值为：<class 'celery.concurrency.prefork.TaskPool'>。
代码如下：
def setup_defaults(self, concurrency=None, loglevel='WARN', logfile=None,
task_events=None, pool=None, consumer_cls=None,
timer_cls=None, timer_precision=None,
autoscaler_cls=None,
pool_putlocks=None,
pool_restarts=None,
optimization=None, O=None, # O maps to -O=fair
statedb=None,
time_limit=None,
soft_time_limit=None,
scheduler=None,
pool_cls=None, # XXX use pool
state_db=None, # XXX use statedb
task_time_limit=None, # XXX use time_limit
task_soft_time_limit=None, # XXX use soft_time_limit
scheduler_cls=None, # XXX use scheduler
schedule_filename=None,
max_tasks_per_child=None,
prefetch_multiplier=None, disable_rate_limits=None,
worker_lost_wait=None,
max_memory_per_child=None, **_kw):
either = self.app.either # 从配置文件中获取，如果获取不到就使用给定的默认值
self.loglevel = loglevel # 设置日志等级
self.logfile = logfile # 设置日志文件
self.concurrency = either('worker_concurrency', concurrency) # 设置worker的工作进程数
self.task_events = either('worker_send_task_events', task_events) # 设置时间
self.pool_cls = either('worker_pool', pool, pool_cls) # 连接池设置
self.consumer_cls = either('worker_consumer', consumer_cls) # 消费类设置
self.timer_cls = either('worker_timer', timer_cls) # 时间类设置
self.timer_precision = either(
'worker_timer_precision', timer_precision,
)
self.optimization = optimization or O # 优先级设置
self.autoscaler_cls = either('worker_autoscaler', autoscaler_cls)
self.pool_putlocks = either('worker_pool_putlocks', pool_putlocks)
self.pool_restarts = either('worker_pool_restarts', pool_restarts)
self.statedb = either('worker_state_db', statedb, state_db) # 执行结果存储
self.schedule_filename = either(
'beat_schedule_filename', schedule_filename,
) # 定时任务调度设置
self.scheduler = either('beat_scheduler', scheduler, scheduler_cls) # 获取调度类
self.time_limit = either(
'task_time_limit', time_limit, task_time_limit) # 获取限制时间值
self.soft_time_limit = either(
'task_soft_time_limit', soft_time_limit, task_soft_time_limit,
)
self.max_tasks_per_child = either(
'worker_max_tasks_per_child', max_tasks_per_child,
) # 设置每个子进程最大处理任务的个数
self.max_memory_per_child = either(
'worker_max_memory_per_child', max_memory_per_child,
) # 设置每个子进程最大内存值
self.prefetch_multiplier = int(either(
'worker_prefetch_multiplier', prefetch_multiplier,
))
self.disable_rate_limits = either(
'worker_disable_rate_limits', disable_rate_limits,
)
self.worker_lost_wait = either('worker_lost_wait', worker_lost_wait)
2.3 setup_instance in worker
执行完成后，会继续执行 self.setup_instance方法来建立实例。
def setup_instance(self, queues=None, ready_callback=None, pidfile=None,
include=None, use_eventloop=None, exclude_queues=None,
**kwargs):
self.pidfile = pidfile # pidfile
self.setup_queues(queues, exclude_queues) # 指定相关的消费与不消费队列
self.setup_includes(str_to_list(include)) # 获取所有的task任务
# Set default concurrency
if not self.concurrency: # 如果没有设置默认值
try:
self.concurrency = cpu_count() # 设置进程数与cpu的个数相同
except NotImplementedError:
self.concurrency = 2 # 如果获取失败则默认为2
# Options
self.loglevel = mlevel(self.loglevel) # 设置日志等级
self.ready_callback = ready_callback or self.on_consumer_ready # 设置回调函数
# this connection won't establish, only used for params
self._conninfo = self.app.connection_for_read()
self.use_eventloop = (
self.should_use_eventloop() if use_eventloop is None
else use_eventloop
) # 获取eventloop类型
self.options = kwargs
signals.worker_init.send(sender=self) # 发送信号
# Initialize bootsteps
self.pool_cls = _concurrency.get_implementation(self.pool_cls) # 获取缓冲池类
self.steps = [] # 需要执行的步骤
self.on_init_blueprint()
self.blueprint = self.Blueprint(
steps=self.app.steps['worker'],
on_start=self.on_start,
on_close=self.on_close,
on_stopped=self.on_stopped,
) # 初始化blueprint
self.blueprint.apply(self, **kwargs) # 调用初始化完成的blueprint类的apply方法
其中setup_queues和setup_includes所做的工作如下，
def setup_queues(self, include, exclude=None):
include = str_to_list(include)
exclude = str_to_list(exclude)
try:
self.app.amqp.queues.select(include) # 添加队列消费
except KeyError as exc:
raise ImproperlyConfigured(
SELECT_UNKNOWN_QUEUE.strip().format(include, exc))
try:
self.app.amqp.queues.deselect(exclude) # 不消费指定的队列中的任务
except KeyError as exc:
raise ImproperlyConfigured(
DESELECT_UNKNOWN_QUEUE.strip().format(exclude, exc))
if self.app.conf.worker_direct:
self.app.amqp.queues.select_add(worker_direct(self.hostname)) # 添加消费的队列
def setup_includes(self, includes):
# Update celery_include to have all known task modules, so that we
# ensure all task modules are imported in case an execv happens.
prev = tuple(self.app.conf.include) # 获取配置文件中的task
if includes:
prev += tuple(includes)
[self.app.loader.import_task_module(m) for m in includes] # 将task添加到loader的task中
self.include = includes
task_modules = {task.__class__.__module__
for task in values(self.app.tasks)} # 获取已经注册的任务
self.app.conf.include = tuple(set(prev) | task_modules) # 去重后重新设置include
2.3.1 setup_queues
self.app.amqp.queues.select(include)会设置queues。
堆栈如下：
__init__, amqp.py:59
Queues, amqp.py:259
queues, amqp.py:572
__get__, objects.py:43
setup_queues, worker.py:172
setup_instance, worker.py:106
__init__, worker.py:99
worker, worker.py:326
caller, base.py:132
new_func, decorators.py:21
invoke, core.py:610
invoke, core.py:1066
invoke, core.py:1259
main, core.py:782
start, base.py:358
worker_main, base.py:374
代码位于：celery/app/amqp.py
class Queues(dict):
"""Queue name⇒ declaration mapping.
"""
#: If set, this is a subset of queues to consume from.
#: The rest of the queues are then used for routing only.
_consume_from = None
def __init__(self, queues=None, default_exchange=None,
create_missing=True, autoexchange=None,
max_priority=None, default_routing_key=None):
dict.__init__(self)
self.aliases = WeakValueDictionary()
self.default_exchange = default_exchange
self.default_routing_key = default_routing_key
self.create_missing = create_missing
self.autoexchange = Exchange if autoexchange is None else autoexchange
self.max_priority = max_priority
if queues is not None and not isinstance(queues, Mapping):
queues = {q.name: q for q in queues}
queues = queues or {}
for name, q in queues.items():
self.add(q) if isinstance(q, Queue) else self.add_compat(name, **q)
所以目前如下：
+----------------------+
+----------+ | @cached_property |
| User | | Worker |
+----+-----+ +---> | |
| | | |
| worker_main | | Worker application |
| | | celery/app/base.py |
v | +----------+-----------+
+---------+------------+ | |
| Celery | | |
| | | |
| Celery application | | v
| celery/app/base.py | | +--------------+--------------+ +---> app.loader.init_worker
| | | | class Worker(WorkController)| |
+---------+------------+ | | | |
| | | +--------> setup_defaults
| celery.main | | Worker as a program | |
| | | celery/apps/worker.py | |
v | +-----------------------------+ +---> setup_instance +-----> setup_queues +------> app.amqp.queues
+---------+------------+ |
| @click.pass_context | |
| celery | |
| | |
| | |
| CeleryCommand | |
| celery/bin/celery.py | |
| | |
+---------+------------+ |
| |
| |
| |
v |
+----------+------------+ |
| @click.pass_context | |
| worker | |
| | |
| | |
| WorkerCommand | |
| celery/bin/worker.py | |
+-----------+-----------+ |
| |
+-----------------+
手机如图：
2.4 Blueprint
下面就要建立 Worker 内部的各个子模块。
worker初始化过程中，其内部各个子模块的执行顺序是由一个BluePrint类定义，并且根据各个模块之间的依赖进行排序（实际上把这种依赖关系组织成了一个 DAG）执行，此时执行的操作就是加载blueprint类中的default_steps。
具体逻辑是：
- self.claim_steps方法功能是获取定义的steps
- _finalize_steps 获取step依赖，并进行拓扑排序，返回按依赖关系排序的steps。
- 根据依赖，返回依赖排序后step。
- 当step生成之后，就开始调用来生成组件。
- apply函数最后，返回worker。当所有的类初始化完成后，此时就是一个worker就初始化完成。
代码如下：celery/apps/worker.py。此时的Blueprint类定义在Worker里面。
class Blueprint(bootsteps.Blueprint):
"""Worker bootstep blueprint."""
name = 'Worker'
default_steps = {
'celery.worker.components:Hub',
'celery.worker.components:Pool',
'celery.worker.components:Beat',
'celery.worker.components:Timer',
'celery.worker.components:StateDB',
'celery.worker.components:Consumer',
'celery.worker.autoscale:WorkerComponent',
}
celery Worker 中各个模块定为step，具体如下：
Timer：用于执行定时任务的 Timer，和 Consumer 那里的 timer 不同；
Hub：Eventloop 的封装对象；
Pool：构造各种执行池（线程/进程/协程）；
Autoscaler：用于自动增长或者 pool 中工作单元；
StateDB：持久化 worker 重启区间的数据（只是重启）；
Autoreloader：用于自动加载修改过的代码；
Beat：创建 Beat 进程，不过是以子进程的形式运行（不同于命令行中以 beat 参数运行）；
celery 中定义依赖关系主要依靠了几个类属性 requires，label，conditon，last，比如Hub依赖Timer，Consumer最后执行。
class Hub(bootsteps.StartStopStep):
"""Worker starts the event loop."""
requires = (Timer,)
class Consumer(bootsteps.StartStopStep):
"""Bootstep starting the Consumer blueprint."""
last = True
2.5 Blueprint基类
apply调用的是基类代码。其基类位于celery/bootsteps.py。
class Blueprint:
"""Blueprint containing bootsteps that can be applied to objects.
"""
GraphFormatter = StepFormatter
name = None
state = None
started = 0
default_steps = set()
state_to_name = {
0: 'initializing',
RUN: 'running',
CLOSE: 'closing',
TERMINATE: 'terminating',
}
def __init__(self, steps=None, name=None,
on_start=None, on_close=None, on_stopped=None):
self.name = name or self.name or qualname(type(self))
self.types = set(steps or []) | set(self.default_steps)
self.on_start = on_start
self.on_close = on_close
self.on_stopped = on_stopped
self.shutdown_complete = Event()
self.steps = {}
apply代码如下，这个是在WorkController初始化时执行的。
def apply(self, parent, **kwargs):
"""Apply the steps in this blueprint to an object.
This will apply the ``__init__`` and ``include`` methods
of each step, with the object as argument::
step = Step(obj)
...
step.include(obj)
For :class:`StartStopStep` the services created
will also be added to the objects ``steps`` attribute.
"""
self._debug('Preparing bootsteps.')
order = self.order = [] # 用于存放依序排列的需要执行的模块类
steps = self.steps = self.claim_steps() # 获得定义的step，生成{step.name, step}结构
self._debug('Building graph...')
for S in self._finalize_steps(steps): # 进行依赖排序后，返回对应的step
step = S(parent, **kwargs) # 获得实例化的step
steps[step.name] = step # 已step.name为key,step实例为val
order.append(step) # 添加到order列表中
self._debug('New boot order: {%s}',
', '.join(s.alias for s in self.order))
for step in order: # 遍历order列表
step.include(parent) # 各个step依次执行include函数，执行create函数
return self
2.5.1 获取定义的steps
其中self.claim_steps方法功能是获取定义的steps，具体如下：
def claim_steps(self):
return dict(self.load_step(step) for step in self.types)# 导入types中的类，并返回已名称和类对应的k:v字典
def load_step(self, step):
step = symbol_by_name(step)
return step.name, step
其中self.types可以在初始化的时候传入。
def __init__(self, steps=None, name=None,
on_start=None, on_close=None, on_stopped=None):
self.name = name or self.name or qualname(type(self))
self.types = set(steps or []) | set(self.default_steps)
self.on_start = on_start
self.on_close = on_close
self.on_stopped = on_stopped
self.shutdown_complete = Event()
self.steps = {}
在Blueprint初始化时没有传入steps，所以此时types就是default_steps属性，该值就是WorkController类中的Blueprint类的default_steps值。
default_steps = {
'celery.worker.components:Hub',
'celery.worker.components:Pool',
'celery.worker.components:Beat',
'celery.worker.components:Timer',
'celery.worker.components:StateDB',
'celery.worker.components:Consumer',
'celery.worker.autoscale:WorkerComponent',
}
2.5.2 _finalize_steps
_finalize_steps作用为： 获取step依赖，并进行拓扑排序，返回按依赖关系排序的steps。
其中重点分析一下self._finalize_steps(steps)函数，排序操作主要在此函数中完成。
def _finalize_steps(self, steps):
last = self._find_last() # 找到last=True的step，上文也提到Consumer类
self._firstpass(steps) # 将step及所需要的依赖加入到steps列表，并获取依赖
it = ((C, C.requires) for C in values(steps)) # 数据结构((a,[b,c]),(b,[e]))
G = self.graph = DependencyGraph( # 依赖图模型初始化，添加定点和边界
it, formatter=self.GraphFormatter(root=last),
)
if last: # last最终执行模块，依赖所有模块执行完成后执行，所以为所有模块添加last到step边界
for obj in G:
if obj != last:
G.add_edge(last, obj)
try:
return G.topsort() # 进行拓扑运算，获得排好序的steps列表
except KeyError as exc:
raise KeyError('unknown bootstep: %s' % exc)
首先定义了一个拓扑类DependencyGraph，根据依赖关系添加定点和边结构：
- 顶点就是各个step。
- 边结构就是依赖step生成的列表。
- 结构为{step1:[step2,step3]}。
下面我们一起看一下DependencyGraph类中的初始化操作
@python_2_unicode_compatible
class DependencyGraph(object):
def __init__(self, it=None, formatter=None):
self.formatter = formatter or GraphFormatter()
self.adjacent = {} # 存储图形结构
if it is not None:
self.update(it)
def update(self, it):
tups = list(it)
for obj, _ in tups:
self.add_arc(obj)
for obj, deps in tups:
for dep in deps:
self.add_edge(obj, dep)
def add_arc(self, obj): # 添加定点
self.adjacent.setdefault(obj, [])
def add_edge(self, A, B): # 添加边界
self[A].append(B)
图形结构已经生成了各个模块之间的依赖关系图，主要依靠celery自己实现的一套DAG，依靠拓扑排序方法，得到执行顺序。拓扑排序就是实现依赖排序，生成模块的执行顺序，然后按照循序执行模块。
2.5.3 返回依赖排序后step
这部分作用就是根据依赖，返回依赖排序后step。
此时由于这几个类中存在相互依赖的执行，比如Hub类，
class Hub(bootsteps.StartStopStep):
"""Worker starts the event loop."""
requires = (Timer,)
Hub类就依赖于Timer类，所以_finalize_steps的工作就是将被依赖的类先导入。
此时继续分析到order列表，该列表就是所有依赖顺序解决完成后的各个类的列表，并且这些steps类都是直接继承或间接继承自bootsteps.Step。
@with_metaclass(StepType)
class Step(object):
...
该类使用了元类，继续查看StepType。
class StepType(type):
"""Meta-class for steps."""
name = None
requires = None
def __new__(cls, name, bases, attrs):
module = attrs.get('__module__') # 获取__module__属性
qname = '{0}.{1}'.format(module, name) if module else name # 如果获取到了__module__就设置qname为模块.name的形式，否则就设置成name
attrs.update(
__qualname__=qname,
name=attrs.get('name') or qname,
) # 将要新建的类中的属性,name和__qualname__更新为所给的类型
return super(StepType, cls).__new__(cls, name, bases, attrs)
def __str__(self):
return bytes_if_py2(self.name)
def __repr__(self):
return bytes_if_py2('step:{0.name}{{{0.requires!r}}}'.format(self))
这里使用了有关Python元类编程的相关知识，通过在新建该类实例的时候控制相关属性的值，从而达到控制类的相关属性的目的。此时会调用Step的include方法。
def _should_include(self, parent):
if self.include_if(parent):
return True, self.create(parent)
return False, None
def include(self, parent):
return self._should_include(parent)[0]
如果继承的是StartStopStep，则调用的include方法如下，
def include(self, parent):
inc, ret = self._should_include(parent)
if inc:
self.obj = ret
parent.steps.append(self)
return inc
排序之后，变量数据举例如下：
order = {list: 7}
0 = {Timer} <step: Timer>
1 = {Hub}
2 = {Pool} <step: Pool>
3 = {WorkerComponent} <step: Autoscaler>
4 = {StateDB} <step: StateDB>
5 = {Beat} <step: Beat>
6 = {Consumer} <step: Consumer>
__len__ = {int} 7
steps = {dict: 7}
'celery.worker.components.Timer' = {Timer} <step: Timer>
'celery.worker.components.Pool' = {Pool} <step: Pool>
'celery.worker.components.Consumer' = {Consumer} <step: Consumer>
'celery.worker.autoscale.WorkerComponent' = {WorkerComponent} <step: Autoscaler>
'celery.worker.components.StateDB' = {StateDB} <step: StateDB>
'celery.worker.components.Hub' = {Hub} <step: Hub>
'celery.worker.components.Beat' = {Beat} <step: Beat>
__len__ = {int} 7
2.5.4 生成组件
当step生成之后，就开始调用来生成组件。
for step in order:
step.include(parent)
对于每个组件，会继续调用。各个step依次执行include函数，执行create函数
def include(self, parent):
return self._should_include(parent)[0]
如果需要调用，就建立组件，比如self为timer，parent是worker实例
<class 'celery.worker.worker.WorkController'>。
代码如下：
def _should_include(self, parent):
if self.include_if(parent):
return True, self.create(parent)
return False, None
def include_if(self, w):
return w.use_eventloop
以timer为例，
class Timer(bootsteps.Step):
"""Timer bootstep."""
def create(self, w):
if w.use_eventloop:
# does not use dedicated timer thread.
w.timer = _Timer(max_interval=10.0)
其堆栈如下：
create, components.py:36
_should_include, bootsteps.py:335
include, bootsteps.py:339
apply, bootsteps.py:211
setup_instance, worker.py:139
__init__, worker.py:99
worker, worker.py:326
caller, base.py:132
new_func, decorators.py:21
invoke, core.py:610
invoke, core.py:1066
invoke, core.py:1259
main, core.py:782
start, base.py:358
worker_main, base.py:374
<module>, myTest.py:26
2.5.5 返回worker
apply函数最后，返回worker。当所有的类初始化完成后，此时就是一个worker就初始化完成。
def apply(self, parent, **kwargs):
"""Apply the steps in this blueprint to an object.
This will apply the ``__init__`` and ``include`` methods
of each step, with the object as argument::
step = Step(obj)
...
step.include(obj)
For :class:`StartStopStep` the services created
will also be added to the objects ``steps`` attribute.
"""
self._debug('Preparing bootsteps.')
order = self.order = []
steps = self.steps = self.claim_steps()
self._debug('Building graph...')
for S in self._finalize_steps(steps):
step = S(parent, **kwargs)
steps[step.name] = step
order.append(step)
self._debug('New boot order: {%s}',
', '.join(s.alias for s in self.order))
for step in order:
step.include(parent)
return self
self如下：
self = {Blueprint} <celery.worker.worker.WorkController.Blueprint object at 0x7fcd3ad33d30>
GraphFormatter = {type} <class 'celery.bootsteps.StepFormatter'>
alias = {str} 'Worker'
default_steps = {set: 7} {'celery.worker.components:Hub', 'celery.worker.components:Consumer', 'celery.worker.components:Beat', 'celery.worker.autoscale:WorkerComponent', 'celery.worker.components:StateDB', 'celery.worker.components:Timer', 'celery.worker.components:Pool'}
graph = {DependencyGraph: 7} celery.worker.autoscale.WorkerComponent(3)\n celery.worker.components.Pool(2)\n celery.worker.components.Hub(1)\n celery.worker.components.Timer(0)\ncelery.worker.components.StateDB(0)\ncelery.worker.components.Hub(1)\n celery.worker.components.Timer(0)\ncelery.worker.components.Consumer(12)\n celery.worker.autoscale.WorkerComponent(3)\n celery.worker.components.Pool(2)\n celery.worker.components.Hub(1)\n celery.worker.components.Timer(0)\n celery.worker.components.StateDB(0)\n celery.worker.components.Hub(1)\n celery.worker.components.Timer(0)\n celery.worker.components.Beat(0)\n celery.worker.components.Timer(0)\n celery.worker.components.Pool(2)\n celery.worker.components.Hub(1)\n celery.worker.components.Timer(0)\ncelery.worker.components.Beat(0)\ncelery.worker.components.Timer(0)\ncelery.worker.components.Pool(2)\n celery.worker.components.Hub(1)\n celery.worker.co...
name = {str} 'Worker'
order = {list: 7} [<step: Timer>, <step: Hub>, <step: Pool>, <step: Autoscaler>, <step: StateDB>, <step: Beat>, <step: Consumer>]
shutdown_complete = {Event} <threading.Event object at 0x7fcd3ad33b38>
started = {int} 0
state = {NoneType} None
state_to_name = {dict: 4} {0: 'initializing', 1: 'running', 2: 'closing', 3: 'terminating'}
steps = {dict: 7} {'celery.worker.autoscale.WorkerComponent': <step: Autoscaler>, 'celery.worker.components.StateDB': <step: StateDB>, 'celery.worker.components.Hub': <step: Hub>, 'celery.worker.components.Consumer': <step: Consumer>, 'celery.worker.components.Beat': <step: Beat>, 'celery.worker.components.Timer': <step: Timer>, 'celery.worker.components.Pool': <step: Pool>}
types = {set: 7} {'celery.worker.autoscale:WorkerComponent', 'celery.worker.components:StateDB', 'celery.worker.components:Hub', 'celery.worker.components:Consumer', 'celery.worker.components:Beat', 'celery.worker.components:Timer', 'celery.worker.components:Pool'}
此时如下：
+----------------------+
+----------+ | @cached_property |
| User | | Worker |
+----+-----+ +---> | |
| | | |
| worker_main | | Worker application |
| | | celery/app/base.py |
v | +----------+-----------+
+---------+------------+ | |
| Celery | | |
| | | |
| Celery application | | v
| celery/app/base.py | | +--------------+--------------+ +---> app.loader.init_worker
| | | | class Worker(WorkController)| |
+---------+------------+ | | | |
| | | +--------> setup_defaults
| celery.main | | Worker as a program | |
| | | celery/apps/worker.py | |
v | +-----------------------------+ +---> setup_instance +-----> setup_queues +------> app.amqp.queues
+---------+------------+ | +
| @click.pass_context | | |
| celery | | +------------------------------+
| | | | apply
| | | |
| CeleryCommand | | v
| celery/bin/celery.py | |
| | | +-------------------------------------+ +---------------------+ +---> claim_steps
+---------+------------+ | | class Blueprint(bootsteps.Blueprint)| | class Blueprint | |
| | | +------>-+ | +-------> _finalize_steps
| | | | | | |
| | | celery/apps/worker.py | | celery/bootsteps.py | | +--> Timer
v | +-------------------------------------+ +---------------------+ +---> include +--->+
+----------+------------+ | +--> Hub
| @click.pass_context | | |
| worker | | +--> Pool
| | | |
| | | +--> ......
| WorkerCommand | | |
| celery/bin/worker.py | | +--> Consumer
+-----------+-----------+ |
| 1 app.Worker |
+-----------------+
手机如下：
0x3 start in worker 命令
此时初始化完成后就执行到了celery/bin/worker.py中，开始执行 worker。
worker.start()
具体回忆之前代码下：
def worker(ctx, hostname=None, pool_cls=None, app=None, uid=None, gid=None,
loglevel=None, logfile=None, pidfile=None, statedb=None,
**kwargs):
"""Start worker instance.
"""
app = ctx.obj.app
if kwargs.get('detach', False):
return detach(...)
worker = app.Worker(...)
worker.start() #我们回到这里
return worker.exitcode
3.1 start in Worker as a program
此时调用的方法就是：celery/worker/worker.py。可以看到，直接就是调用 blueprint的 start函数，就是启动 blueprint 之中各个组件。
def start(self):
try:
self.blueprint.start(self) # 此时调用blueprint的start方法
except WorkerTerminate:
self.terminate()
except Exception as exc:
self.stop(exitcode=EX_FAILURE)
except SystemExit as exc:
self.stop(exitcode=exc.code)
except KeyboardInterrupt:
self.stop(exitcode=EX_FAILURE)
3.2 start in blueprint
代码是：celery/bootsteps.py
此时，parent.steps就是在step.include中添加到该数组中，parent.steps目前值为[Hub,Pool,Consumer]，此时调用了worker的on_start方法。本例本例如下:
parent.steps = {list: 3}
0 = {Hub} <step: Hub>
1 = {Pool} <step: Pool>
2 = {Consumer} <step: Consumer>
具体 start 代码如下：
class Blueprint:
"""Blueprint containing bootsteps that can be applied to objects.
"""
def start(self, parent):
self.state = RUN
if self.on_start:
self.on_start()
for i, step in enumerate(s for s in parent.steps if s is not None):
self.started = i + 1
step.start(parent)
3.2.1 回调 on_start in Worker
blueprint首先回调 on_start in Worker。
此时代码是/celery/apps/worker.py
具体是：
- 设置app；
- 用父类的on_start；
- 打印启动信息；
- 注册相应的信号处理handler；
- 做相关配置比如重定向；
class Worker(WorkController):
"""Worker as a program."""
def on_start(self):
app = self.app # 设置app
WorkController.on_start(self) # 调用父类的on_start
# this signal can be used to, for example, change queues after
# the -Q option has been applied.
signals.celeryd_after_setup.send(
sender=self.hostname, instance=self, conf=app.conf,
)
if self.purge:
self.purge_messages()
if not self.quiet:
self.emit_banner() # 打印启动信息
self.set_process_status('-active-')
self.install_platform_tweaks(self) # 注册相应的信号处理handler
if not self._custom_logging and self.redirect_stdouts:
app.log.redirect_stdouts(self.redirect_stdouts_level)
# TODO: Remove the following code in Celery 6.0
# This qualifies as a hack for issue #6366.
warn_deprecated = True
config_source = app._config_source
if isinstance(config_source, str):
# Don't raise the warning when the settings originate from
# django.conf:settings
warn_deprecated = config_source.lower() not in [
'django.conf:settings',
]
3.2.2 基类on_start
此时代码是/celery/apps/worker.py。
def on_start(self):
app = self.app
WorkController.on_start(self)
WorkController代码在：celery/worker/worker.py。
父类的on_start就是创建pid文件。
def on_start(self):
if self.pidfile:
self.pidlock = create_pidlock(self.pidfile)
3.2.3 信号处理handler
其中注册相关的信号处理handler的函数如下，
def install_platform_tweaks(self, worker):
"""Install platform specific tweaks and workarounds."""
if self.app.IS_macOS:
self.macOS_proxy_detection_workaround()
# Install signal handler so SIGHUP restarts the worker.
if not self._isatty:
# only install HUP handler if detached from terminal,
# so closing the terminal window doesn't restart the worker
# into the background.
if self.app.IS_macOS:
# macOS can't exec from a process using threads.
# See https://github.com/celery/celery/issues#issue/152
install_HUP_not_supported_handler(worker)
else:
install_worker_restart_handler(worker) # 注册重启的信号 SIGHUP
install_worker_term_handler(worker)
install_worker_term_hard_handler(worker)
install_worker_int_handler(worker)
install_cry_handler() # SIGUSR1 信号处理函数
install_rdb_handler() # SIGUSR2 信号处理函数
单独分析一下重启的函数，
def _reload_current_worker():
platforms.close_open_fds([
sys.__stdin__, sys.__stdout__, sys.__stderr__,
])# 关闭已经打开的文件描述符
os.execv(sys.executable, [sys.executable] + sys.argv)# 重新加载该程序
以及：
def restart_worker_sig_handler(*args):
"""Signal handler restarting the current python program."""
set_in_sighandler(True)
safe_say('Restarting celery worker ({0})'.format(' '.join(sys.argv)))
import atexit
atexit.register(_reload_current_worker) # 注册程序退出时执行的函数
from celery.worker import state
state.should_stop = EX_OK # 设置状态
platforms.signals[sig] = restart_worker_sig_handler
其中的platforms.signals类设置了setitem方法，
def __setitem__(self, name, handler):
"""Install signal handler.
Does nothing if the current platform has no support for signals,
or the specified signal in particular.
"""
try:
_signal.signal(self.signum(name), handler)
except (AttributeError, ValueError):
pass
此时就将相应的handler设置到了运行程序中，_signal就是导入的signal库。
3.2.4 逐次调用step
然后，blueprint逐次调用step的start。
此时，parent.steps就是在step.include中添加到该数组中，parent.steps目前值为[Hub,Pool,Consumer]，此时调用了worker的on_start方法，
parent.steps = {list: 3}
0 = {Hub} <step: Hub>
1 = {Pool} <step: Pool>
2 = {Consumer} <step: Consumer>
此时就继续执行Blueprint的start方法，
依次遍历parent.steps的方法中，依次遍历Hub,Pool,Consumer，调用step的start方法。
def start(self, parent):
self.state = RUN # 设置当前运行状态
if self.on_start: # 如果初始化是传入了该方法就执行该方法
self.on_start()
for i, step in enumerate(s for s in parent.steps if s is not None): # 依次遍历step并调用step的start方法
self._debug('Starting %s', step.alias)
self.started = i + 1
step.start(parent)
logger.debug('^-- substep ok')
3.2.4.1 Hub
由于Hub重写了start方法，该方法什么都不执行，
def start(self, w):
pass
3.2.4.2 Pool
继续调用Pool方法，此时会调用到StartStopStep,此时的obj就是调用create方法返回的对象，此时obj为pool实例,
def start(self, parent):
if self.obj:
return self.obj.start()
具体我们后文讲解。
3.2.4.3 Consumer
继续调用Consumer的start方法，
def start(self):
blueprint = self.blueprint
while blueprint.state != CLOSE: # 判断当前状态是否是关闭
maybe_shutdown() # 通过标志判断是否应该关闭
if self.restart_count: # 如果设置了重启次数
try:
self._restart_state.step() # 重置
except RestartFreqExceeded as exc:
crit('Frequent restarts detected: %r', exc, exc_info=1)
sleep(1)
self.restart_count += 1 # 次数加1
try:
blueprint.start(self) # 调用开始方法
except self.connection_errors as exc:
# If we're not retrying connections, no need to catch
# connection errors
if not self.app.conf.broker_connection_retry:
raise
if isinstance(exc, OSError) and exc.errno == errno.EMFILE:
raise # Too many open files
maybe_shutdown()
if blueprint.state != CLOSE: # 如果状态不是关闭状态
if self.connection:
self.on_connection_error_after_connected(exc)
else:
self.on_connection_error_before_connected(exc)
self.on_close()
blueprint.restart(self) # 调用重启方法
Consumer也有blueprint，具体step如下：
- Connection：管理和 broker 的 Connection 连接
- Events：用于发送监控事件
- Agent：
cellactor - Mingle：不同 worker 之间同步状态用的
- Tasks：启动消息 Consumer
- Gossip：消费来自其他 worker 的事件
- Heart：发送心跳事件（consumer 的心跳）
- Control：远程命令管理服务
此时又进入到了blueprint的start方法，该blueprint的steps值是由Consumer在初始化的时候传入的。
传入的steps是Agent,Connection,Evloop,Control,Events,Gossip,Heart,Mingle,Tasks类的实例，然后根据调用最后添加到parent.steps中的实例就是[Connection,Events,Heart,Mingle,Tasks,Control,Gossip,Evloop]，此时就依次调用这些实例的start方法。
首先分析一下Connection的start方法，
def start(self, c):
c.connection = c.connect()
info('Connected to %s', c.connection.as_uri())
就是调用了consumer的connect()函数，
def connect(self):
"""Establish the broker connection.
Retries establishing the connection if the
:setting:`broker_connection_retry` setting is enabled
"""
conn = self.app.connection_for_read(heartbeat=self.amqheartbeat) # 心跳
# Callback called for each retry while the connection
# can't be established.
def _error_handler(exc, interval, next_step=CONNECTION_RETRY_STEP):
if getattr(conn, 'alt', None) and interval == 0:
next_step = CONNECTION_FAILOVER
error(CONNECTION_ERROR, conn.as_uri(), exc,
next_step.format(when=humanize_seconds(interval, 'in', ' ')))
# remember that the connection is lazy, it won't establish
# until needed.
if not self.app.conf.broker_connection_retry: # 如果retry禁止
# retry disabled, just call connect directly.
conn.connect() # 直接连接
return conn # 返回conn
conn = conn.ensure_connection(
_error_handler, self.app.conf.broker_connection_max_retries,
callback=maybe_shutdown,
) # 确保连接上
if self.hub:
conn.transport.register_with_event_loop(conn.connection, self.hub) # 使用异步调用
return conn # 返回conn
此时就建立了连接。
继续分析Task的start方法，
def start(self, c):
"""Start task consumer."""
c.update_strategies() # 更新已知的任务
# - RabbitMQ 3.3 completely redefines how basic_qos works..
# This will detect if the new qos smenatics is in effect,
# and if so make sure the 'apply_global' flag is set on qos updates.
qos_global = not c.connection.qos_semantics_matches_spec
# set initial prefetch count
c.connection.default_channel.basic_qos(
0, c.initial_prefetch_count, qos_global,
) # 设置计数
c.task_consumer = c.app.amqp.TaskConsumer(
c.connection, on_decode_error=c.on_decode_error,
) # 开始消费
def set_prefetch_count(prefetch_count):
return c.task_consumer.qos(
prefetch_count=prefetch_count,
apply_global=qos_global,
)
c.qos = QoS(set_prefetch_count, c.initial_prefetch_count) # 设置计数
此时就开启了对应的任务消费，开启消费后我们继续分析一下loop的开启，
def start(self, c):
self.patch_all(c)
c.loop(*c.loop_args())
此时就是调用了consumer中的loop函数，该loop函数就是位于celery/worker/loops.py中的asyncloop函数。
stack如下：
asynloop, loops.py:43
start, consumer.py:592
start, bootsteps.py:116
start, consumer.py:311
start, bootsteps.py:365
start, bootsteps.py:116
start, worker.py:203
worker, worker.py:327
caller, base.py:132
new_func, decorators.py:21
invoke, core.py:610
invoke, core.py:1066
invoke, core.py:1259
main, core.py:782
start, base.py:358
worker_main, base.py:374
asyncloop函数如下：
def asynloop(obj, connection, consumer, blueprint, hub, qos,
heartbeat, clock, hbrate=2.0):
"""Non-blocking event loop.""" # 其中obj就是consumer实例
RUN = bootsteps.RUN # 获取运行状态
update_qos = qos.update
errors = connection.connection_errors
on_task_received = obj.create_task_handler() # 创建任务处理头
_enable_amqheartbeats(hub.timer, connection, rate=hbrate) # 定时发送心跳包
consumer.on_message = on_task_received # 设置消费的on_message为on_task_received
consumer.consume() # 开始消费
obj.on_ready() # 调用回调函数
obj.controller.register_with_event_loop(hub) # 向所有生成的blueprint中的实例注册hub
obj.register_with_event_loop(hub)
# did_start_ok will verify that pool processes were able to start,
# but this will only work the first time we start, as
# maxtasksperchild will mess up metrics.
if not obj.restart_count and not obj.pool.did_start_ok():
raise WorkerLostError('Could not start worker processes')
# consumer.consume() may have prefetched up to our
# limit - drain an event so we're in a clean state
# prior to starting our event loop.
if connection.transport.driver_type == 'amqp':
hub.call_soon(_quick_drain, connection)
# FIXME: Use loop.run_forever
# Tried and works, but no time to test properly before release.
hub.propagate_errors = errors
loop = hub.create_loop() # 创建loop,本质是一个生成器
try:
while blueprint.state == RUN and obj.connection: # 检查是否在运行并且连接是否有
# shutdown if signal handlers told us to.
should_stop, should_terminate = (
state.should_stop, state.should_terminate,
)
# False == EX_OK, so must use is not False
if should_stop is not None and should_stop is not False:
raise WorkerShutdown(should_stop)
elif should_terminate is not None and should_stop is not False:
raise WorkerTerminate(should_terminate)
# We only update QoS when there's no more messages to read.
# This groups together qos calls, and makes sure that remote
# control commands will be prioritized over task messages.
if qos.prev != qos.value:
update_qos()
try:
next(loop) # 循环下一个
except StopIteration:
loop = hub.create_loop()
finally:
try:
hub.reset()
except Exception as exc: # pylint: disable=broad-except
logger.exception(
'Error cleaning up after event loop: %r', exc)
至此，异步Loop就开启了，然后就开始了服务端的事件等待处理，worker流程就启动完毕。
0x4 本文总结
本文主要讲述了worker的启动流程。celery默认启动的就是以进程的方式启动任务，然后异步IO处理消费端的事件，至此worker的启动流程概述已分析完毕。
+----------------------+
+----------+ | @cached_property |
| User | | Worker |
+----+-----+ +---> | |
| | | |
| worker_main | | Worker application |
| | | celery/app/base.py |
v | +----------+-----------+
+---------+------------+ | |
| Celery | | |
| | | |
| Celery application | | v
| celery/app/base.py | | +--------------+--------------+ +---> app.loader.init_worker
| | | | class Worker(WorkController)| |
+---------+------------+ | | | |
| | | +--------> setup_defaults
| celery.main | | Worker as a program | |
| | | celery/apps/worker.py | |
v | +-----------------------------+ +---> setup_instance +-----> setup_queues +------> app.amqp.queues
+---------+------------+ | +
| @click.pass_context | | |
| celery | | +------------------------------+
| | | | apply
| | | |
| CeleryCommand | | v
| celery/bin/celery.py | |
| | | +-------------------------------------+ +---------------------+ +---> claim_steps
+---------+------------+ | | class Blueprint(bootsteps.Blueprint)| apply | class Blueprint | |
| | | +------>-+ | +-------> _finalize_steps
| | | | | | |
| | | celery/apps/worker.py | | celery/bootsteps.py | | +--> Timer
v | +-----+---------------+---------------+ +---------------------+ +---> include +--->+
+----------+------------+ | ^ | +--> Hub
| @click.pass_context | | | | start |
| worker | | | | +--> Pool
| | | | | +----> WorkController.on_start |
| | | | | | +--> ......
| WorkerCommand | | | +-------------> Hub. start / Pool.start / Task.start |
| celery/bin/worker.py | | | | +--> Consumer
+-----------+-----------+ | | +----> Evloop.start +-----> asynloop
| 1 app.Worker | |
+-----------------+ |
| |
| 2 blueprint.start |
+------------------------->+
手机如下：
0xFF 参考
celery源码分析-worker初始化分析（下）https://blog.csdn.net/qq_33339479/article/details/80958059