def __init__(self, processor, zmq_socket_type=zmq.DEALER, context=None, pool_size=5, mode_ppworker=False, service=None):
# 1. 获取zeromq context, 以及 events
self.context = context or Context.get_instance()
self.events = Events(zmq_socket_type, self.context, mode_ppworker=mode_ppworker, service=service)
# 2. 设置: mode_ppworker
if self.events.mode_ppworker:
self.liveness = HEARTBEAT_LIVENESS
self.interval = INTERVAL_INIT
self.heartbeat_at = time.time() + HEARTBEAT_INTERVAL
self.processor = processor # thrift processor
self.proto_factory_input = TBinaryProtocolFactory()
self.proto_factory_output = TUtf8StrBinaryProtocolFactory()
# 4. gevent
self.task_pool = gevent.pool.Pool(size=pool_size)
self.acceptor_task = None
self.events.create_worker_socket()
self.endpoint = None
# 5. 程序退出控制
self.alive = True
def __init__(self, endpoint, sock_type = zmq.DEALER, ctx = None, timeout=5, profile=False): # zmq.DEALER
"""
如果采用了 local proxy,
:param endpoint:
:param sock_type:
:param ctx:
:param service:
:return:
"""
self.pid = os.getpid()
self._context = ctx or Context.get_instance() # 获取zeromq context
self._events = Events(sock_type, self._context)
self._events.setsockopt(zmq.IDENTITY, "client-%4d" % os.getpid())
self._endpoint = endpoint
self._wbuf = StringIO()
self._rbuf = StringIO()
self.service = None
self.timeout = timeout * 1000 # seconds --> milli-seconds
self.timeout_in_second = timeout
self.seqNum = 0
self.profile = profile
class Server(object):
def __init__(self, processor, zmq_socket_type=zmq.DEALER, context=None, pool_size=5, mode_ppworker=False, service=None):
# 1. 获取zeromq context, 以及 events
self.context = context or Context.get_instance()
self.events = Events(zmq_socket_type, self.context, mode_ppworker=mode_ppworker, service=service)
# 2. 设置: mode_ppworker
if self.events.mode_ppworker:
self.liveness = HEARTBEAT_LIVENESS
self.interval = INTERVAL_INIT
self.heartbeat_at = time.time() + HEARTBEAT_INTERVAL
self.processor = processor # thrift processor
self.proto_factory_input = TBinaryProtocolFactory()
self.proto_factory_output = TUtf8StrBinaryProtocolFactory()
# 4. gevent
self.task_pool = gevent.pool.Pool(size=pool_size)
self.acceptor_task = None
self.events.create_worker_socket()
self.endpoint = None
# 5. 程序退出控制
self.alive = True
def get_heartbeat_msg(self):
# 协议: byte0(动作) + byte1(版本) + byte2(并发度)
# 启动时发一个消息，之后间隔一段时间发送一个消息
#
return PPP_HEARTBEAT + chr(0) + chr(self.task_pool.free_count())
def get_ready_msg(self):
# 协议: byte0(动作) + byte1(版本) + byte2(并发度)
# 启动时发一个消息，之后间隔一段时间发送一个消息
#
return PPP_READY + chr(0) + chr(self.task_pool.free_count())
# 通过_events来connect, bind服务
def connect(self, endpoint, resolve=True):
self.endpoint = endpoint
return self.events.connect(endpoint, resolve)
def bind(self, endpoint, resolve=True):
self.endpoint = endpoint
return self.events.bind(endpoint, resolve)
def close(self):
self.events.close()
self.stop()
def handle_request(self, event):
# 1. 将zeromq的消息转换成为 thrift的 protocols
trans_input = TMemoryBuffer(event.msg)
trans_output = TMemoryBuffer()
proto_input = self.proto_factory_input.getProtocol(trans_input)
proto_output = self.proto_factory_output.getProtocol(trans_output)
# 2. 交给processor来处理
try:
self.processor.process(proto_input, proto_output)
# 3. 将thirft的结果转换成为 zeromq 格式的数据
msg = trans_output.getvalue()
# print "Return Msg: ", msg, event.id
self.events.emit(msg, event.id)
except Exception as e:
# 如何出现了异常该如何处理呢
# 程序不能挂
print "Exception: ", e
def _acceptor(self):
# run
# ---> _acceptor
# ---> _handle_request
#
#
# server的工作模式:
# 1. Demo服务器可以简单地启动一个ZeroRpcServer, 然后也不用考虑网络 io的一点点时间开销
# 2. 线上服务器, ZeroRpcServer之前添加了一个queue或load balance, 因此网络io的时间也可以忽略
#
last_queue_time = time.time()
last_event_time = time.time()
start = True
while True:
if self.events.mode_ppworker:
# 注意: events现在只从 input获取信息, 如果 handle_request 因为什么原因，导致数据没有返回，而代码却堵在这里了。
# poll_event的参数最小为1，否则就成为0, 则无限等待; 也不能设置为None
event = self.events.poll_event(1) # 1ms(为什么呢?)
now = time.time()
if event:
last_queue_time = now
last_event_time = now
if self.liveness == 0 or start:
start = False
print Fore.GREEN, "Queue back to life", Fore.RESET
# print "Event: ", event
if len(event.msg) == 1 and event.msg == PPP_HEARTBEAT:
self.liveness = HEARTBEAT_LIVENESS
else:
# 正常的RPC数据
self.liveness = HEARTBEAT_LIVENESS
self.task_pool.spawn(self.handle_request, event)
self.interval = INTERVAL_INIT
else:
# timeout(太长时间没有回应)
if self.alive and (now - last_queue_time >= HEARTBEAT_INTERVAL):
last_queue_time = time.time()
now = time.time()
self.liveness -= 1
if self.liveness == 0:
print Fore.RED, "ID: ", self.events.identity, "Queue died, Begin to sleep for: ", self.interval, Fore.RESET
# 反正都没啥事了，等待就等待
gevent.sleep(self.interval)
if self.interval < INTERVAL_MAX:
self.interval *= 2
self.liveness = HEARTBEAT_LIVENESS
start = True
# 重新注册(恢复正常状态)
assert self.endpoint
print Fore.RED, "ID: ", self.events.identity, "Reconnection Queue", Fore.RESET
self.events.reconnect(self.get_ready_msg())
# 如果要死了，并且1s内没有新的任务, 并且task_pool为空，则自杀
if not self.alive and (now - last_event_time) > 1.0 and self.task_pool.free_count() == self.task_pool.size:
print Fore.CYAN, "<<<< Suicide Worker Gracefully", Fore.RESET
break
# 自己也需要发送消息给queue(自己活着的时候才发送hb)
elif now >= self.heartbeat_at and self.alive:
print "Send Hb Msg...", self.events.identity
self.events.emit(self.get_heartbeat_msg(), None)
self.heartbeat_at = time.time() + HEARTBEAT_INTERVAL
if not self.alive:
self.events.emit(PPP_STOP, None)
else:
event = self.events.poll_event(1)
if event:
self.task_pool.spawn(self.handle_request, event)
def run(self):
import gevent.monkey
gevent.monkey.patch_socket()
# 0. 注册信号(控制运维)
self.init_signal()
# 1 ppworker在启动的时候，会告知 lb, 我准备好了，以及支持的并发度
if self.events.mode_ppworker:
print "Send Hb Msg..."
self.events.emit(self.get_ready_msg(), None)
# 2. 监听数据
self.acceptor_task = gevent.spawn(self._acceptor)
# 3. 等待结束
try:
self.acceptor_task.get()
finally:
self.stop()
self.task_pool.join(raise_error=True)
def stop(self):
if self.acceptor_task is not None:
self.acceptor_task.kill()
self.acceptor_task = None
def init_signal(self):
def handle_int(*_):
self.alive = False
def handle_term(*_):
# 主动退出
self.alive = False
# 2/15
signal.signal(signal.SIGINT, handle_int)
signal.signal(signal.SIGTERM, handle_term)
print Fore.RED, "To graceful stop current worker plz. use:", Fore.RESET
print Fore.GREEN, ("kill -15 %s" % os.getpid()), Fore.RESET
class TZmqTransport(TTransportBase, CReadableTransport):
def close(self):
self._events.close()
# 通过_events来connect, bind服务
def connect(self, endpoint, resolve=True):
return self._events.connect(endpoint, resolve)
def bind(self, endpoint, resolve=True):
return self._events.bind(endpoint, resolve)
def __init__(self, endpoint, sock_type = zmq.DEALER, ctx = None, timeout=5, profile=False): # zmq.DEALER
"""
如果采用了 local proxy,
:param endpoint:
:param sock_type:
:param ctx:
:param service:
:return:
"""
self.pid = os.getpid()
self._context = ctx or Context.get_instance() # 获取zeromq context
self._events = Events(sock_type, self._context)
self._events.setsockopt(zmq.IDENTITY, "client-%4d" % os.getpid())
self._endpoint = endpoint
self._wbuf = StringIO()
self._rbuf = StringIO()
self.service = None
self.timeout = timeout * 1000 # seconds --> milli-seconds
self.timeout_in_second = timeout
self.seqNum = 0
self.profile = profile
def open(self):
self.connect(self._endpoint)
def read(self, size):
ret = self._rbuf.read(size)
if len(ret) != 0:
return ret
# buf中的数据处理完毕了，说明一个message应该处理完毕了
self._read_message()
return self._rbuf.read(size)
def get_seq_num(self, event_ids):
for id in event_ids:
if id:
return int(id)
else:
continue
return None
def _read_message(self):
t = time.time() + self.timeout_in_second
event = self._events.poll_event(self.timeout)
while event:
# print "EventId: ", event.id
seqNum = self.get_seq_num(event.id)
if self.profile:
print event.id
if not event._id[0]:
event._id = event._id[1:]
print "Seq: ", event.id[0], "--->",
start = float(event.id[2])
for i in range(4, len(event.id), 2):
t = float(event.id[i])
print "%.4fs" % (t - start),
start = t
print "%.4fs" % (time.time() - start)
# 等待下一个event
# 1. seqNum一致, 读取成功；
# 2. 读取到之前出现的event, 然后继续等待，
# 3. 或者放弃
if seqNum != self.seqNum:
print "SeqNum Not match: ", seqNum, " exp: ", self.seqNum
left = t - time.time()
if left > 0:
# Case 2
event = self._events.poll_event(int(left * 1000))
print "------- Next: ", left
else:
# Case 3
event = None # 没有读取到有效的Event
break
else:
# Case 1
break
if not event:
raise TimeoutException()
else:
self._rbuf = StringIO(event.msg)
def write(self, buf):
self._wbuf.write(buf)
def flush(self):
# 有时进程被fork了，zeromq socket被多个进程共享，则结果会出现混乱
if os.getpid() != self.pid:
raise UnexpectedForkException()
msg = self._wbuf.getvalue()
self._wbuf = StringIO()
# 范围: 1 ~ 10000 (防止seqNum过长)
self.seqNum += 1
if self.seqNum > SEQ_NUM_MAX:
self.seqNum = SEQ_NUM_MIN
# 将Thrift转换成为zeromq
if self.service:
# <service, '', msg>
if self.profile:
self._events.emit(msg, [self.service, "", str(self.seqNum), "", "%.4f" % time.time()]) # client似乎没有id
else:
self._events.emit(msg, [self.service, "", str(self.seqNum)]) # client似乎没有id
else:
# <msg>
if self.profile:
self._events.emit(msg, [str(self.seqNum), "", "%.4f" % time.time()]) # client似乎没有id
else:
self._events.emit(msg, [str(self.seqNum)]) # client似乎没有id
# Implement the CReadableTransport interface.
@property
def cstringio_buf(self):
return self._rbuf
# NOTE: This will probably not actually work.
def cstringio_refill(self, prefix, reqlen):
while len(prefix) < reqlen:
self.read_message()
prefix += self._rbuf.getvalue()
self._rbuf = StringIO(prefix)
return self._rbuf
