ZeroMQ's variety of socket semantics is great, and it does a good job abstracting away the messy details of TCP. You can build rather complicated systems on top of it much easier and faster than you could using raw sockets. The implementation itself has some rough spots, such as the multi threading. I really like the elegance of nanomsg, but it never really got the same polish that ZeroMQ did. I wouldn't expect the same raw throughput as a hand-tuned TCP based protocol, but it's generally more than fast enough for most use cases.
I believe the original developer passed away a few years ago?
And he was passing away due to cancer came to HN for one of his last discussions. He also mentioned he was broke and a PayPal account for donations for his family was mentioned.
nkurz on Apr 18, 2016 [–]
Thanks. Does anything come to mind that you should have done more of (or instead)?
PieterH on Apr 18, 2016 [–]
Probably more frequent checkups, to be honest. Though it's probably moot. "
Pieter Hintjens, whom you are referring to, was actually not so much involved in the development of the core ZeroMQ library, though he certainly contributed. His main role was building a community and he did develop a high level C binding czmq.
Martin Sustrik was actually what could be referred to as the lead engineer for the initial ZeroMQ project, though he left it quite some time ago.
I have to say that I have never really had a positive impression of zeromq. I feel like it is particularly unsuccessful and that seems to be for some pretty good reasons.
Zeromq is a bag of tools and it's not really clear to me how a novice would put them together safely. This is important for a message system because everyone has to start somewhere and it really shouldn't be "step 1: build your broker with our subroutines". Is that broker really going to be any good?
Secondly, for many years zeromq was a kind of monty python style 'judean people's front vs people's front of judea' community. They split from the AMQP standardisation process I agree had clearly gone wildly wrong and then...split and then...split again.
Right as far as I know people doing messaging are using a) RabbitMQ if they're deploying it themselves b) AWS of GCP's thing or c) some proprietary thing which has no name recognition among the public (who here has heard of AMPS?) d) Kafka. No one is using zeromq as far as I know! Look forward to anyone sharing their positive experiences using zeromq on big projects eg in banks.
As someone who's written more than my fair share of messaging code using a bunch of different middleware and frameworks, I've never really understood where ZeroMQ is supposed to fit. It's high-level enough that you have to learn something completely different than plain sockets (and have more stuff to debug when it breaks), while also being low-level enough that it can't provide many of the features I'd expect out of a message-oriented middleware-- no message broker, no anonymous pub-sub, no reliable transport, etc. unless you go and build them yourself (which they provide documentation and patterns for, but they're not things you get for free like you would with other frameworks).
"Bag of tools" is about right. Where RabbitMQ/AMQP or MQTT or DDS can, depending on exactly what I want to do, provide me an end-to-end solution where, with a bit of configuration and network management, throw in a message and expect it to go where I want it to, ZeroMQ hands me a pile of pieces with which I can, presumably, build a thing that will do the same thing. That's not something I've ever found that I need-- either one of the premade solutions fills my needs, or I go and write something myself and skip ZeroMQ.
ZeroMQ is used heavily in the backend of Jupyter and allows for a huge variety of non-Python kernels. It's brokerless, which sets it apart from all the other things you listed here.
Forgive me as you seem to be aware of many of the details of zmq’s history but is there a chance you’re missing something? You are comparing a thin, semantic-enforcing layer just above TCP to actual message queues/servers featuring persistence to disk, etc.
We’ve used ZeroMQ in production as an alternative to HTTP message passing and it was great for that. We would build something like RabbitMQ atop of ZMQ, the only similarity between the two being the name, really.
More to the point, while you could build a theoretical HareMQ atop of ZMQ, you can also build a hundred different products/services that aren’t a RabbitMQ alternative out of it, too. Apples and oranges.
I could well be missing something! I get that zeromq is not quite equivalent to rabbit or Google pubsub but you do presumably still want the same kinds of patterns: pubsub, queues and maybe a bit of routing?
So it is an alternative to (eg) Google pubsub and I think can be compared with it. Is there any case where it would be used for a different /purpose/ than the others?
I agree. The basic interface design seems really neat but it's so easy to run into horrible problems with it. For one it doesn't really deal with errors very well: once you've thrown a message into it there's basically no mechanism to know if something goes wrong with it (and it certainly used to just straight up assert in many, many error cases, which sure made it hard to use the 'reliable central broker' mechanism they advocate in the docs). The implementation is also really brittle: it's basically trivial to cause all kinds of misbehaviour if the other end does not act as expected. The problem is there isn't really many good alternatives that are actually similar. Other 'message queues' are basically big databases, not a library. I feel there's room for something with the same philosophy but better execution.
The only times I've ever reached for zeromq are when I needed sockets and didn't want to deal with the hassle of programming sockets - it worked great then.
I'd happily use it as the glue between two processes that I control, but would grab one of the other tech you mentioned for anything exposed on the boundaries of my app.
In 2011 I was a novice and got that bag of tools dropped in my lap, and I did struggle with how to put them together, but safely was not. problem. It was considerably better for a novice to play with those tools than to have to do it at the socket level.
We are using 0mq for monitoring of electric vehicles fleet. Millions of messages every day (hundreds of GB), zero problems with bad internet connection (common in moving vehicles). Every sent message is received regardless of connectivity issues. Just a little implementation annoyances here and there but it is serving us well since 2015.
How does 0mq achieve that? For all I have read and tried to use of it, 0mq explicitly offers no message delivery guarantees, instead relying on your own implementation to achieve them. So, I don't doubt that your system works like you described, but I think you may be overvaluing 0mq's contribution to it.
Yes, I have read the guide pretty thoroughly. What all those things have in common is that they are your responsibility to implement - all that 0mq does is reconnect the socket, and queue up messages. Anything else that could give you reliability is either TCP's problem or yours. That includes timeouts, retransmit, ACKs, keepalives, heartbeats, rate control, message storage, everything that ensures reliability - with 0mq you get to implement it. Lazy pirate indeed.
We implemented KISS state machine based on TCP timeouts and server response.
Pseudocode:
Send msg to server
if socket timeout:
reinit socket and send again
Recv from server
if recv.msg != 'ok':
send again
else
remove msg from storage
One message has from few to few hundreds Kilobytes. Messages are stored and sent as custom binary data for better compression (better that protobuffs).
ZMQ takes complexity of raw tcp sockets away and this is what we needed it for. As much and as little.
This leaves higher application code to handle duplicated messages. It also does not handle cases where the client connects later than the server send, but would still like to receive the server messages. It also doesn't handle rate control (overwhelming the network).
These may all be fine for your application, and then 0mq is an excellent fit. But I think that for most applications, having some kind of pre-implemebted solution to all of these problems, even with slightly more overhead, is extremely valuable.
My only complaint with the guide so far was that it was historically presented as a single huge DokuWiki page with ambiguous heading structures (probably because the page was made of several other pages, each having top-level headings by its own). The current version looks much cleaner!
100% agree.
I think what stood out to me the most was that it did explain a lot of the decision making. I feel like often other docs don't explain the decisions that did go into their software at all
This is also what sets good code comments and code level API documentation apart - explaining the reasoning and intention for both implementation and for intended use. Essentially the why and how, not that what...
I looked at zeromq, but it seemed like it had very bad debug-ability and visibility into internal operations. Their FAQ [0] says things like:
> How do I determine how many messages are in queue?
> This isn't possible. [...] rather than provide incorrect information the library avoids providing any view into this data.
> How can I retrieve a list of all connected peers?
> This is not supported.
Those kinds of decisions are very scary for production environment. We've had to debug network connectivity problems before, and it just possible what all the interfaces Linux kernel provides. I cannot imagine doing it with library which hides stuff from you on purpose.
> At any given time a message may be in the ZeroMQ sender queue, the sender's kernel buffer, on the wire, in the receiver's kernel buffer or in the receiver's ZeroMQ receiver queue. Furthermore, a ZeroMQ socket can bind and/or connect to many peers. Each peer may have different performance characteristics and therefore a different queue depth. Any "queue depth" number is almost certainly wrong
The library doesn't hide stuff from you, it just refuses to show you information that might as well be a random number.
If you want to know how many peers are connected you use heartbeats instead of just blindly trusting that 1 tcp socket = 1 available peer.
The reasonings assumes I don't know what's going on. This approach is precisely why I never tried ZMQ.
How do you know my circumstances? Maybe this system has tuned rmem_max/wmem_max so I don't care about kernel buffers? Maybe my programs have problems with multi-gigabyte queues? Why are you assuming stuff for me?
The right answer of course is you don't expose a single "queue size" value, but you don't hide it either. Instead, expose something like "list of sockets: for each, here is an fd and a size of userspace queue". I'll then query the kernel myself to get sender's and receiver's kernel buffer size and so on.
Life is hard enough without libraries hiding stuff from you!
But how do I implement backpressure in zeromq? I've used it in production and it was great when it worked, but it was scary because its queue of messages could get arbitrarily large. I think I even saw an OoM when a service went down and something just queued indefinitely.
Edit: just checked the FAQ and apparently it will now block or drop messages. I sort of wish it would explicitly error instead of block, so as not to require the caller to queue in a thread or something.
To be fair, when you use regular sockets you also don't get much visibility into what your OS is doing, by default. If zeromq provided the information it would definitely be used by app developers misguided to how they are supposed to use it, exactly.
One thing that I didn't like while using it was not being able to get logs out of it. So, you don't know if the issue is from zeromq or app itself. Most of the time, it was app itself but also found a few a bugs in Zeromq which took quite time to find it.
If I remember correctly, they’re written by the original zeromq author. He has a blog post about how he realized that C++ was not ideal for the architecture he was using, and decided to make a clean break.
Problem is - nng is also hardly being updated. I was really hopeful for a C library that would by on par with ZeroMQ but every time a new project comes along that needs some sort of message bus and I have the time to evaluate the technology, ZeroMQ simply wins.
His arguments are valid [1] but I still don't understand why that means dropping C++ entirely. Sure, C++ isn't a perfect superset of C, but you could mostly write C-style code in C++, right?
I've seen this sentiment before from C enthusiasts. Maybe it makes sense on the Linux kernel, where you don't want to keep telling contributors not to use the most common C++ features.
But I don't think it makes sense on very small projects, and maybe I just draw the dividing line for where it stops working at a different size of project than he did.
[1] Exceptions suck and constructors suck and C++ would be better if it was just Rust with no borrow-checker.
I've implemented bindings for ZeroMQ with Tokio in rust (https://github.com/cetra3/tmq). I found ZeroMQ great for interop with Python/Java if you just want some simple cross-process broadcast/task management.
Software is a huge domain. It’s just one of those things you’ll need when you need.
Some places you might use it: real-time financial data streams, multi-agent worker task/management, distributed database syncing. It’s basically a networking lib that makes common patterns easier. E.g send this message to everyone, list connected peers, handle these messages this way if the peer can’t take the message, and so on. They support doing this in a bunch of languages over a bunch of communication protocols.
Now all the Async/coroutine/ipc verbiage is mostly just because synchronization is a communication problem which is what zeroMQ happens to solve as well.
ØMQ promises being "socket on steroids" but I used it quite a bit and it's not so beneficial.
It hides the scalability problem of having many TCP sockets without solving it (e.g. pub-sub over TCP).
It does not solve the ahead-of-line blocking issue.
It requires reading a lot of documentation and doing experiments to find out what is the right combination of " fan-out, pub-sub, task distribution, and request-reply" that you have to use.
Plus a lot of "magic" happening in the library involves queues that you have no control over. You get applications that hang and fill buffers instead of failing early.
I suspect this is the reason why it never became a new layer in the TCP/IP stack
zmq is "a better socket()", but unlike sockets it's pretty hard to figure out what's going on. E.g. most (all?) socket types don't need the other side to be immediately available, it'll buffer a couple thousand or so messages and start blocking if it doesn't appear. From the app's point of view, everything seems fine until the queues are full (which you can't introspect, btw.) when suddenly everything starts blocking.
It does bring some nice things to the table, but using it has always felt at least a little bit "weird" to me. The docs are a little weird, and there are random caveats sprinkled around the library (e.g. you don't need to synchronize the startup of dependent services like described above, unless you were using certain transports in older versions and so on).
It used to be marketed somewhat towards developing multi-threaded applications, but it was always rather unclear what exactly the advantage of using, for example, the comparatively slow PUSH/PULL socket (ipc:// are pipes) over a MPMC (or as needed) queue ought to be.
It is also marketed towards resilient distributed apps, but, perhaps I was holding it wrong, but it'd always behave much like TCP and stuff like REQ/REP would pretty much always block forever if the peer goes away, while stuff like PUSH/PULL isn't reliable when peers go away, so for a distributed app these can't really be used anyway. So for this use case it always seemed rather too low-level and its guarantees too weak / non-applicable.
> It is also marketed towards resilient distributed apps, but, perhaps I was holding it wrong, but it'd always behave much like TCP and stuff like REQ/REP would pretty much always block forever if the peer goes away, while stuff like PUSH/PULL isn't reliable when peers go away,
I had exactly the same (disappointing) experience.
zmq appears to be a layer that tries to hide some of the most important details that the socket layer exposes. socket programming and the state machines that tend to go along with it are some of the least difficult parts of systems programming to do well, it is well documented, well understood and there are ample tools to help measure ground truth. it was always weird to me that zmq had such a strong reputation; it basically hides what would otherwise be a few hundred lines of library code (including a framer).
You can think of ZeroMQ as legos for building message queue-based messaging systems. ZeroMQ gives you different pieces - pushers, pullers, publishers, subscribers, routers, etc. and it's up to you to then hook them up in different ways to create the system you want. The guide gives you some different ideas of how to hook up the different pieces for common use cases, but there's nothing stopping you from building a custom one.
ZeroMQ is not an out-of-the-box MQ solution. It takes some work to understand and build a functional MQ solution with ZeroMQ. If you want something that works out of the box I recommend RabbitMQ.
For internal machine-to-machine communications you can replace your full blown HTTP server + JSON inefficencies with a simple ZMQ request/reply + binary serialisation format.
It also allows one machine to broadcast a message to a 'topic' that other machines listen for.
Many other things. It's not perfect, but it Just Works.
”ZeroMQ (also spelled ØMQ, 0MQ or ZMQ) is a high-performance asynchronous messaging library, aimed at use in distributed or concurrent applications. It provides a message queue, but unlike message-oriented middleware, a ZeroMQ system can run without a dedicated message broker.”
I read it as a low-level abstraction for when you want to build a custom messaging infrastructure.
I believe the 'problem' that it solves is universal network transparency between pieces of anything. It's a library that can be used from basically any language that has a C FFI on any system with a C compiler to do IPC with anything else supporting same using a uniform and relatively simple API.
It's an alternative to writing your own complex networking code for any kind of distributed communications.
A lot of people these days might use something like flink or spark to build a pipeline whereas I'd prefer to write a number of small servers with various fan out or fan in operations, broadcast style operations, request/reply semantics etc where appropriate
Makes way more sense and much simpler to my mind and allows much better control over performance characteristics.
I am sure that the developers behind zmq are(or were) far more intelligent and better programmers than I will ever be, but I have to say that as a user I would never recommend it to anyone for the following reasons:
1. A lot of gotchas and surprising behavior. For example this:
> To be more explicit: after calling zmq_send the socket may become readable (and vice versa) without triggering a read event on the file descriptor.
2. The library will assert internally without actually giving you any clues what did you do wrong as a user.
I would rather spend some extra time in development by using pure sockets rather than the stress of debugging problems, arising from the above, later.
I am actually using it in production as an easy mechanism to add non critical rpc calls between services. The only thing I can say about it is that zeromq hasn't given me a reason to look for a replacement. Yet.
I believe that zeromsg does not support true zero-copy messaging which might be a reason to look at nanomsg or nng if you need every last ounce of performance.
The fact that it just works, and very easy to add endpoints, made it a no-brainer. I might not leave it alone with my kids but it works great as an RPC pipe.
Ø is a vowel and a letter used in the Danish, Norwegian, Faroese, and Southern Sami languages. (Wikipedia)
It’s not an unused letter that you can snap up as a cool way to write zero. People will read it, some will recognize it and assume it has meaning, because it has. That’s how letters work.
There are zeroes with dashes through them, but if you look closely they almost always take care to keep the dash inside the (non-)circle zero to make the difference clearer.
I realize most people in the world won’t care about this, but with Northern European eyes, this looks stupid.
The ∅ symbol (U+2205 in Unicode), even though it's inspired by, is not the same as the letter Ø (U+00D8 in Unicode), and has been, for almost a century already, the "correct" way to represent an empty set in mathematical notation (alongside {}).
That is the meaning everyone that does not primarily use the Northern European alphabet (aka most people in the world) assume it has.
Maybe suppress your instinctive outrage next time, or if you really need to be angry at something, at least direct it at the right place: the dead mathematician who proposed it [0]
I copypasted the character from their website, which is why I know it’s not U+2205 being used. They’re also easy to tell apart.
Maybe suppress your instinctive correction next time until you know that the person you’re correcting is in fact mistaken. Or just use a nicer tone to begin with.
> The Ø in ZeroMQ is all about tradeoffs. On the one hand this strange name lowers ZeroMQ’s visibility on Google and Twitter. On the other hand it annoys the heck out of some Danish folk who write us things like “ØMG røtfl”, and “Ø is not a funny looking zero!” and “Rødgrød med fløde!”, which is apparently an insult that means “may your neighbours be the direct descendants of Grendel!” Seems like a fair trade.
Because I'm sure more than one person is wondering, the literal translation of "rødgrød med fløde" is "red porridge with heavy cream", not an insult but a traditional danish dessert.
I saw that, and it doesn’t do anything for me personally than paint the authors as petulant. Clearly “some danish folk” are the obvious other here and they themselves can’t be bothered because they don’t read the character. If they want to refer to the mathematical character as mentioned in other comments, why not use that instead of one from an actual alphabet? They’re free to do whatever they want but anyone familiar with the character is also free to see them as being not a little US-centric.
Ø - The 28. Letter in the Norwegian alphabet, pronounced as the beginning of the sound ‘uuuuh’ (like when someone is making a sound when thinking). It’s a join between O and E, and was at first written as Ө. The same sound in Swedish is Ö
> The Ø in ZeroMQ is all about tradeoffs. On the one hand this strange name lowers ZeroMQ’s visibility on Google and Twitter. On the other hand it annoys the heck out of some Danish folk who write us things like “ØMG røtfl”, and “Ø is not a funny looking zero!” and “Rødgrød med fløde!”, which is apparently an insult that means “may your neighbours be the direct descendants of Grendel!” Seems like a fair trade.
I've used ZeroMQ lightly as a socket-replacement and for simple IPC and I'd say the key thing it solved was connection establishment. It turns out that opening a two way connection between two apps on the same network without a clear starting order can go wrong in many ways and become quite messy to get right. With ZeroMQ it more or less just magically worked and it allowed me to focus on the actual problem I was working on.
The wide range of language support was also nice. It allowed me to test the interface using a Python notebook which came in handy for debugging.
If you have ever written non trivial TCP sockets code you would notice for many purposes that using this saves you lots of very non trivial socket handling intricacies.
For me 0mq works internally, and not for interfacing outside a closed system. I think that might be the source of some confusion
When people say that they mean it's a replacement for the use of sockets in your program, not that it replaces sockets ... Across the technological landscape? Within the tech stack?
If your program is the only user of sockets then okay, "use of sockets in your program" might be a valid assumption. But if you're interfacing with external components using sockets/whatever, using ZMQ will not solve anything. That's why I don't see it as a replacement/competing standard...
In 2015, in one project one "clever" co-worker decided to use "zeromq" in project where is the guarantee of delivery was must. I think that because he wanted to attach this technology as fancy stuff in his cv. Later he left and we had to rewrite message sending and converted to basic http.
(Yes to the earlier post - had a similar experience with an early metrics system for Twilio in 2012-2013)
The 1st issue is that ZMQ implementations generally have an IO thread handling all socket IO. REQ/REP patterns tend to work out just fine, because the client is going to wait for a reply from the server. But things like PUSH/PULL become...harder to follow. You can push things in the client, but how do you know when they've flushed out of the IO thread and made it to the server? Those sorts of things matter when shutting down a process, running integration tests, etc. In contrast, with an HTTP request, you're basically always doing REQ/REP, and so you know when the data has been pushed.
The 2nd issue is that ZMQ implementations tend to be harder to observe / operate than a more common path like HTTP requests. (HTTP requests can go through a load balancer, have standard response codes, use headers for authorization, etc.) For these reasons, I've tended to avoid ZMQ ever after -- it seems like often, you're best off either using REST or GRPC if by HTTP, or raw TCP if it's purely a data push kind of operation (e.g. forwarding structured logs, with framing, to a remote TLS endpoint).
About the same time, I used zeromq in a project for similar reasons:
- I wanted easy message framing
- I didn't know much about networks
- HTTP servers that can be embedded in C++ aren't great, and I didn't know anything about HTTP
- It is nice that you can ignore application startup order if you don't care much about reliability, which I didn't
If I knew then what I know now, I would probably have picked something else. For one, I couldn't make it integrate with my main event loop, so I had it block in another thread, and that caused a problem of inter-thread communication that was nearly as bad as the original problem, for me as an inexperienced noobie.
> I couldn't make it integrate with my main event loop
You can get a file descriptor from ZMQ to use with your own event loop (depending on your event loop of course). I have done this successfully in the past with an epoll (iirc) based event loop.
The documentation is at [1], the ZMQ_FD option. You do have to read the instructions very carefully though.
I believe the original developer passed away a few years ago?