written on Tuesday, June 26, 2012
ZeroMQ is a cool technology. It's advertised as “better sockets” and I can get behind that idea for the most part. Unfortunately the high abstraction also makes it very easy to write applications that become unresponsive. That's because ZeroMQ has an implicit state machine for a very common socket type that can get stuck in the wrong state if the peer disconnects.
If you walk over to the ZeroMQ website you will find an example of the REQ / REP socket pair. It works very similar to TCP sockets. One the one side you bind the socket to an interface and port, on the other side you connect to it. However there is one huge difference between those two.
When talking about TCP sockets the server socket has a way to accept a new client connection. In the mainloop you typically wait until you accept a new connection which then “produces” a new socket. In ZeroMQ there are always only two sockets in the request / response relationship. So how do you send replies back if a request comes? It works because ZeroMQ switches state of the request socket. When the request socket gets a request it memorizes where the request came from and the next sent response goes to that socket.
This example server shows the general pattern that is provided by the documentation. it runs in a loop and receives a request and sends a reply back:
import zmq context = zmq.Context() socket = context.socket(zmq.REP) socket.connect('tcp://*:5555') while 1: # Get request request = socket.recv() # Reply with response socket.send('this is anwer to ' + request)
The client to this is very simple as well. Instead of using REP it uses REQ as socket type, connect, sends a message and waits for one reply:
import zmq context = zmq.Context() socket = context.socket(zmq.REQ) socket.connect('tcp://localhost:5555') socket.send('Hello') message = socket.recv() print 'Received reply: %s' % message
From the same page that introduces REP and REQ sockets you can also find this wonderful paragraph:
Now this looks too simple to be realistic, but a ØMQ socket is what you get when you take a normal TCP socket, inject it with a mix of radioactive isotopes stolen from a secret Soviet atomic research project, bombard it with 1950-era cosmic rays, and put it into the hands of a drug-addled comic book author with a badly-disguised fetish for bulging muscles clad in spandex. Yes, ØMQ sockets are the world-saving superheroes of the networking world.
Now I hope you did not get too much carried away with marketing speak because I certainly were. Unfortunately it seems like there is something terribly wrong with these general pattern.
It's basically this question that shows the problem: what happens if a disconnect happens between the send and the recv? Since ZeroMQ hides disconnects entirely from you, there is no way to be notified if the client or the server went away.
Sadly the socket has an implicit state attached so if a disconnect happens while the socket is in the bound state everything falls apart. And to the best of my knowledge there is no way to force the socket into a different state by hand.
Imagine this piece of code on the server:
import zmq while 1: # Receive a request request = socket.recv() # Reply with a response, but sadly we get an exception raise Exception('Unfortunately this failed') socket.send('this is answer to ' + request)
So before the server was able to reply with a response the server died of an exception. Very unfortunate. If this was a TCP based server you would just restart the server after this incident, the client got a connection reset and would hopefully try again. In ZeroMQ you're now in a worse spot after the server restart. The server's socket is in the “accepting requests” state again but the client never noticed the server to disappear and is still waiting for the reply to come. This however will never happen.
This behavior should probably be obvious but I have never seen a ZeroMQ example that did not suffer from that. The Python bindings make that even worse by providing you with recv_json and send_json methods. What happens if you send invalid JSON data to such a ZeroMQ endpoint? Correct: an exception and most likely a stuck client.
But this is not limited to uncaught exceptions. The authorization flow of salt is taking long enough on our servers that if you restart the salt master in the wrong moment you end up with a stuck minion. Very annoying behavior because it means someone has to SSH into the server restarting the stuck process. If your server management tool gets stuck that's very much the worst case scenario.
With that knowledge, how do you prevent this from happening? Unfortunately there is no easy cure. The only thing I can think of is adding timeouts to all calls. Since ZeroMQ by itself does not really provide simple timeout functionality I recommend using a subclass of the ZeroMQ socket that adds a timeout parameter to all important calls:
from functools import update_wrapper import zmq class Socket(zmq.Socket): def __init__(self, ctx, type, default_timeout=None): zmq.Socket.__init__(self, ctx, type) self.default_timeout = default_timeout def on_timeout(self): return None def _timeout_wrapper(f): def wrapper(self, *args, **kwargs): timeout = kwargs.pop('timeout', self.default_timeout) if timeout is not None: timeout = int(timeout * 1000) poller = zmq.Poller() poller.register(self) if not poller.poll(timeout): return self.on_timeout() return f(self, *args, **kwargs) return update_wrapper(wrapper, f, ('__name__', '__doc__')) for _meth in dir(zmq.Socket): if _meth.startswith(('send', 'recv')): locals()[_meth] = _timeout_wrapper(getattr(zmq.Socket, _meth)) del _meth, _timeout_wrapper
Now instead of calling s.recv() you would call s.recv(timeout=5.0) and if a response does not come back within that 5 second window it will return None and stop blocking. For most applications it probably makes more sense to raise some sort of signalling exception in on_timeout which bubbles up to whatever mainloop is there and handle it there.
This could probably be improved by having a background thread that uses a ZeroMQ socket for heartbeating. That way you could detect disconnects easier. The problem with the general timeout is that some request/replies might take a while to go through and legitimate replies could exceed that 5 second window. Also if you expect a lot of disconnects the 5 seconds might be too much in general.
I was thinking about the problem a little bit and I am not sure if there is a better solution for that particular problem. Disconnect detection is a general problem if there is a high level of abstraction. I was running into a similar problem when I tried using redis lists as a communication layer for a monitoring system. If the application dies it will not have the chance to send a disconnect signal, at least not reliably — a timeout is the only chance you have.
However on the TCP level the operating system is responsible for sending disconnect information and heartbeating which makes this a lot more reliable. Even if your app is stuck for 60 seconds doing nothing, the OS will still not terminate the connection. On top of that is TCP something that multiple components along the route speak. Network equipment in the middle also close your connection if they detect problems on the network.
A userspace level communication protocol like ZeroMQ can't really benefit from all of that. ZeroMQ could provide an API that provides disconnect information if it can get that from the layer it sits on, but part of the ZeroMQ design is that TCP is just one possible implementation and I don't know if it helps to bring disconnect information back.
I suppose the better solution for the problem would be to memorize the state the socket is in in your app and to bring it by hand into the state before the crash/restart. If for instance I could write the application in a way that if I restart it, it restarts exactly at the point where it left off, ZeroMQ's behavior would work fine again.