August 18, 2015
I have been an enthusiast of ZeroMQ for quite some time. If there was an opportunity that required some sockets on steroids, I would not think twice. Ah, how those three messaging pattern were useful (i.e., Request/Reply, Push/Pull, Publish/Subscribe). They had the remarkable trait of being able to drastically streamline the development of distributed systems. And its portability, with wrappers for more than 30 languages, brokerlessness and amazing documentation made it a no-brainer for me to favour ZeroMQ over other messaging alternatives.
Using ZeroMQ in Python with PyZMQ, however, always made me feel like I'm coding in C/C++, which I also love by the way. Unfortunately, that lack of PyZMQ API's, if I may say, “pythonicity”, just felt wrong to me. And by the end of last January, I decided to do something about it. So, that is how Zeroless was born.
My mission was to leverage on PyZMQ to build a more elegant wrapper for ZeroMQ. Something more aligned with the python way of doing things. And, to a certain degree, I have succeeded. However, I have never made a comprehensive effort to publicise Zeroless in any way so, in this post, I hope not only to explain how Zeroless differs from PyZMQ but also to reach a greater audience, that may be as enthusiastic about ZeroMQ as myself. Therefore, without further ado, here goes some of the design decisions I have made for Zeroless:
TCP only for the win
Ok, I know PGM, INPROC and IPC have their use cases. PGM, for instance, provides a Publish/Subscribe specific transport, that scales better than TCP in the Publish/Subscribe use case, as it cut out the ACK flood publishers gets on every new message. There are also some extra reliability, that you also cannot find in TCP. IPC, on the other hand, is a pattern agnostic way of providing more efficient inter-process communication than traditional networking, but is Unix-like only. As for INPROC's particular case, which efficient applicability is being hindered by the Python's GIL, I do not see why bother with it.
Nevertheless, I have a feeling that the vast majority of the users, like myself, are ok with just TCP being supported. Which is precisely what you need when building really horizontally scalable networked services, especially in this time of a renewed vision for SOA, with microservices having a lot of attention. So let us just use TCP and free our minds to think about other matters.
No more contexts
PyZMQ applications require users to create a context, in order to instantiate sockets. Technically, a context serves as a container for all your sockets and usually one of it, per process, is just what you require. As a matter of fact, you could have more, but why bother your runtime with more event loops, for your socket stuff, when one suffices? Moreover, if you are using INPROC as transport, you may also need to share a context for the communication to happen. But again, if INPROC is not that useful in Python as aforementioned, do we really need explicitly manage contexts?
Not at all, so that is why in Zeroless you just have to manage Clients, sockets that connect, and Servers, sockets that bind, without concerning yourself with contexts ;). For instance, in order to instantiate a client you would:
client = Client() # You could use connect_local(port=12345) as well client.connect(ip='127.0.0.1', port=12345)
Similarly, for servers, you would:
server = Server(port=12345) # No need to call bind here
Finally, note that no real connect/bind will occur unless you instantiate a messaging pattern, which are the subject of our next topic.
Like a factory method pattern
One thing I never liked about PyZMQ sockets's instantiation is that we have to call a method called socket, which receives an enum representing the type of the socket. Why don't they just provide a separate method for every socket possible, like as if it were a factory method pattern kind of interface? That would allow a more straightforward experience for developers, that could then rely on their favourite IDE's code complete to quickly understand what kind of sockets and parameters they could set. That enum approach, however, will probably make your users go to the documentation, but solely because of the way the PyZMQ's interface is.
In Zeroless, we fixed that, so you don't need to check the documentation every time you want to instantiate a socket, instead just have a decent code complete support, and you are done. For instance, compare how you would instantiate a publisher socket with Zeroless:
pub = Server(port=12345).pub(topic=b'')
One of the questions you may ask is to whom your clients are connected to. And for that, PyZMQ cannot help you. Unless you manage that list of connections by yourself, you wouldn't be able to get it afterwards. Therefore, in Zeroless, we provide an addresses property, so that you can retrieve all your IP and port pairs as a list of tuples. But that is only for clients of course, as there is no way to know which sockets are connected to your server without building some sort of infrastructure for that yourself.
Subscribe should not be tricky
In terms of interface, the subscribe case is particularly problematic in PyZMQ. One must use the not so intuitive setsockopt() method, in order to define the topics it subscribes to. Like in the following snippet:
socket = context.socket(zmq.SUB) socket.setsockopt(zmq.SUBSCRIBE, b"") # Subscribe to all topics
I believe most new ZeroMQ's user get this wrong at first, as they suppose no topic means you are subscribed to all topics, and keep asking himself/herself why that damn subscriber socket does not receive your published messages.
In Zeroless, we fixed that, so that you don't have to instantiate your socket and set something as essential as a topic, in the subscribe case, via some kind of “obscure” method. Just compare how you would instantiate a subscriber socket with Zeroless:
listen_for_pub = client.sub(topics=[b''])
Generators and high-order functions as first class citizens
PyZMQ sockets tend to use send() and recv() methods for the message exchange part. However, it always felt wrong to me to do stuff like:
while True: data = socket.recv() # do something with data
That is, if Python has built-in support for iterables, or generators if you prefer, why don't we just do something like:
listen_for_push = Server(port=12345).pull() for data in listen_for_push: # do something with data
Way more idiomatic to read incoming messages that way, right? As for sending them, I also followed a different path.
push = client.push() push(b"Msg1")
Therefore, in Zeroless, every time you instantiate a message pattern that is supposed to send messages, use it as a function. Otherwise, treat it as a generator.
Multi-part made easy
In PyZMQ, if you want to send a multipart message, you have to use the recv_multipart() and send_multipart() methods. Methods that, instead of a single message, will deal with a list of them. In Zeroless, I favoured consistency for a quicker and easier learning path. Therefore, there is no difference between the single part and the multipart API.
If you want to send a multipart message, just consider that your send function have a printf like interface, and you are set. So, for instance, if you wish to send an id separated from your message body, you could:
push = client.push() push(b'1', b'OK')
Additionally, if someone sends you a multipart message, your generator will return a tuple with all of its parts. As a result of that, to get the message from the previous example you would need to:
listen_for_push = Server(port=12345).pull() for id, msg in listen_for_push: # do something with id and msg
Although feature parity was never part of my plans, there are still some of PyZMQ's functionalities I would like to provide in Zeroless. Like both poller and devices APIs, for instance. So expect more on the way o/. While that, if you felt compelled to help shape this project, please clone our repository and see our guidelines.