ZeroMQ PUB socket buffers all my out going data when it is connecting

Question

I noticed that a zeromq PUB socket will buffers all outgoing data if it is connecting, for example

import zmq
import time
context = zmq.Context()

# create a PUB socket
pub = context.socket (zmq.PUB)
pub.connect("tcp://127.0.0.1:5566")
# push some message before connected
# they should be dropped
for i in range(5):
    pub.send('a message should not be dropped')

time.sleep(1)

# create a SUB socket
sub = context.socket (zmq.SUB)
sub.bind("tcp://127.0.0.1:5566")
sub.setsockopt(zmq.SUBSCRIBE, "")

time.sleep(1)

# this is the only message we should see in SUB
pub.send('hi')

while True:
    print sub.recv()

The sub binds after those messages, they should be dropped, because PUB should drop messages if no one connected to it. But instead of dropping messages, it buffers all messages.

a message should not be dropped
a message should not be dropped
a message should not be dropped
a message should not be dropped
a message should not be dropped
hi

As you can see, those "a message should not be dropped" are buffered by the socket, once it gets connected, it flush them to SUB socket. If I bind at the PUB socket, and connect at the SUB socket, then it works correctly.

import zmq
import time
context = zmq.Context()

# create a PUB socket
pub = context.socket (zmq.PUB)
pub.bind("tcp://127.0.0.1:5566")
# push some message before connected
# they should be dropped
for i in range(5):
    pub.send('a message should not be dropped')

time.sleep(1)

# create a SUB socket
sub = context.socket (zmq.SUB)
sub.connect("tcp://127.0.0.1:5566")
sub.setsockopt(zmq.SUBSCRIBE, "")

time.sleep(1)

# this is the only message we should see in SUB
pub.send('hi')

while True:
    print repr(sub.recv())

And you can only see the output

'hi'

This kind of strange behavior cause a problem, it buffers all data on a connecting socket, I have two servers, server A publishes data to server B

Server A -- publish --> Server B

It works fine if server B gets online. But what if I start the Server A and do not start Server B?

As the result, the connecting PUB socket on Server A keeps all those data, the memory usage gets higher and higher.

Here is the problem, is this kind of behavior a bug or feature? If it is feature, where can I find a document that mentions this behavior? And how can I stop the connecting PUB socket buffers all data?

Thanks.