Hacker News new | past | comments | ask | show | jobs | submit login
ZeroMQ instead of HTTP, for internal services (augustl.com)
206 points by augustl on July 22, 2013 | hide | past | favorite | 99 comments

To me it looks like a ZeroMQ client can be written in many languages so you don't need to know Clojure to use it. Therefore the article has an extra barrier to entry for understanding the content because more people understand OOP languages like Java than Clojure.

I guess my point is, don't skim the article and think it's not for you because you don't know Clojure. Here's a list of languages it supports ( http://zguide.zeromq.org/page:all#Ask-and-Ye-Shall-Receive ):

C++ | C# | Clojure | CL | Delphi | Erlang | F# | Felix | Go | Haskell | Haxe | Java | Lua | Node.js | Objective-C | Perl | PHP | Python | Q | Racket | Ruby | Scala | Tcl | Ada | Basic | ooc

Here is a more complete list of language bindings taken from http://www.zeromq.org/bindings:_start

Ada, Bash, Basic, C, Chicken Scheme, Common Lisp, C#, C++, D, Delphi, Erlang, F#, Felix, Flex (ActionScript), Go, Guile Scheme, Haskell, Haxe, Java, JavaScript (Node, Flash), Julia, LabVIEW, Lua, Nimrod, Objective-C, Objective Caml, ooc, Perl, PHP, Python, Q, Racket, R, REBOL (2 and 3), Red, Ruby, Scala, Smalltalk, Tcl, XPCOM

OP here, thanks for making that clear. I intentionally tried to keep the Clojure code as generally easy to read as possible, and aside from the big Clojure win (more specifically, functional programming win) towards the end, the code is pretty generic ZeroMQ stuff. I suppose it's not a coincidence that the ZeroMQ docs has the code snippets available in many programming languages :)

I have to confess that I, as a non-Clojure-speaker, found it much less interesting after seeing that your example was in Clojure. Using less common languages as examples makes it feel like it's in that big world of amusing side projects instead of something trying to be mainstream. Not on purpose, it's subconscious.

OTOH, Clojure gets more common as more people use it in examples. Subconsciously, if you keep seeing this strange language every other example code then you start to think that maybe it's not so side-projecty after all.

On the contrary, if I see someone writing in Clojure or Scala it is a good indicator that they can be taken seriously.

That's really not fair, to be honest. The ony way for small languages to break through is to get a lot of exposure so that people start using them.

I wrote an ATS binding which I use in a number of projects: https://github.com/doublec/ats-libzmq

I attempted to use ATS features to make sure resources are cleaned up and msg data usage is statically checked for correct sizes, etc.

Is it now possible to use ZeroMQ in JVM languages without JNI? I used ZeroMQ with Java 2 years ago and had to use JNI to make it work. (Which I did not like).

Yes, you can. The ZeroMQ git repository now contains a pure Java implementation of ZMTP/2.0 which aims to be compatible with libzmq 3.2.2.


I've recently started to migrate my ZeroMQ/JNI to pure JeroMQ and I have a few simple components in production already.

OP here, the examples are using JeroMQ. Actually, that's a lie, I use JZMQ. But the APIs are identical :)

OT, but someone should go update the Golang version to use zerocopy.

In my experience, ZeroMQ is very fragile in RPC-like applications, especially because it tries to abstract away the "connection". This mindset is very appropriate when you're doing multicast (and ZeroMQ rocks when doing multicast), but for unicast, I actually want to detect a disconnection or a connection failure and handle it appropriately before my outgoing buffers are choked to death. So, I'd evaluate other alternatives before settling on ZeroMQ as a transport for internal RPC-type messaging.

If you are fine with having the whole message in memory before parsing (or sending) it (Http is not that bad when it comes to transferring huge documents over the network), writing raw MessagePack document to a regular TCP stream (or tucking it inside a UDP datagram) will do the trick just fine. MessagePack library does support parsing streams -- see e.g. its Python example in its homepage (http://msgpack.org).

Disclosure: I'm just a happy MessagePack (and sometimes ZeroMQ) user. I work on Spyne (http://spyne.io) so I just have experience with some of the most popular protocols out there.

I always enjoy reading discussions of internal application layer communication protocols. If there was one thing I would like to add though or wish for, it would be that people would go one level deeper in their thoughts and discussions. Specifically I'm referring to the transport layer (TCP/IP). I come from the financial services sector and we are heavy users of reliable multicast as an internal transport.

The kinds of solutions, changes, and re-thinking it has allowed us to do with regards to our architecture have just been tremendous. For example, if your system is latency sensitive you can move a lot of your more intensive logging to dedicated nodes. All they have to do is listen on the specific multicast port. This allows one to reduce latency in the critical path because you don't have to spend time logging inputs and outputs which ultimately ends up as some form of disk write or other I/O. Obviously you can't get rid of all of the logging/disk writes that your system must perform to maintain durability and other availability guarantees that are required, but for most things I've noticed that the requirements aren't always that stringent.

What's even cooler about multicast is that there is nothing preventing you from re-implementing your existing point-to-point, sequential, guaranteed message delivery protocols on top of it. A reasonable question to ask would be why would any sane person person want to do this?

Because most of the time you know which sorts of requests are going to result in calls to a manageable and finite number of which services. When using point to point protocols you've essentially painted yourself into the proverbial corner because now somewhere in your architecture you're forced to write fan-ins and fan-outs at the application layer that accumulate the results of queries to various services you utilize. These things are difficult to implement, frequently a source of headache, and difficult to reason about in the event of a failure. Why not use a transport protocol that gives you all of that for free?

The real difficulty here is that in order to properly utilize one-to-many communication internally you really have to have it on your mind since it will require you to structure your software in a very specific way. It's certainly not a silver bullet, and won't cure all that ails you. But now that we've been using it for so long I'm constantly coming up with solutions or ideas that are only possible because we utilize multicast as a transport that would be otherwise too costly or tedious to implement using TCP/IP.

TCP is certainly overkill in my case, since all my RPCs happens on a very stable network with a small and local-ish topology.

Using "fire and forget" multicast for logging and perception sounds interesting. And modelling systems in terms of perception sounds like it's generally a good idea. For truly "perceptive" systems you shouldn't coordinate the observers, since the death of a single observer should be irrelevant. If all your RPCs are idempotent (such as using "PUT" with a client-generated UUID instead of "POST" with no id) this should work well. You could also use vector clocks for eventual consistency. But I suppose one should also make sure to not re-implement dynamo. Or?

What kind of multicast are you using?

ZeroMQ can do multicast using PGM -- http://api.zeromq.org/2-1:zmq-pgm

PGM :)

Ilya Grigorik has been advocating using ZeroMQ with SPDY for modern backend RPC.

See "Building a Modern Web Stack"...

AirBnB Tech Talk: https://www.youtube.com/watch?v=ZxfEcqJ4MOM

Slides: http://www.igvita.com/slides/2012/http2-for-fun-and-profit.p...

Blog: http://www.igvita.com/2012/01/18/building-a-modern-web-stack...

BTW: I posted a ticket to add backend SPDY support to App Engine -- see Ilya's comment: https://code.google.com/p/googleappengine/issues/detail?id=9...

The thing that really annoys me about ZeroMQ is that sockets aren't thread-safe. This isn't such a big problem C, but it caused some really weird and difficult to track down problems in a Haskell program I wrote where I used a PAIR socket, but I wrote and read in different threads. I ended up doing some fairly unsightly socket gymnastics to get around this limitation.

Haskell programs are generally thread-safe by construction. It's very easy to overlook thread-safety in foreign libraries, but this isn't just a problem in Haskell, multi-threaded code is the default is a number of realms.

I now try to avoid ZMQ where I can. Modern libraries should be thread-safe.

Have you found any brokerless mq alternatives? About to build a haskell project and 0mq looks good. I'll keep in mind setting up two sockets, for push and pull. The documentation shows, zeromq contexts being threadsafe, wheras passing off a 0mq socket to another thread requires, putting memory barriers around it.


You should have a look at Nitro.


It's still young, but it grew out of some frustrations with using ZeroMQ in large-scale deployments.


Yes very good. thanks.

ZeroMQ's approach is that of many threads where you don't share any state across them, you just pass messages. I'm not an expert on the subject, so all I can say is that according to the authors of ZeroMQ, this is a good way to proceed, since coordination can take up a lot of resources.

I can only talk from my own experience. In this case, I had two processes talking to each other across a PAIR socket in an asynchronous manner. It was really convenient for me to have one thread waiting on messages on the socket and another taking input from other parts of the process and writing to the socket.

You might argue that in this case having two sockets, a PUSH and a PULL might have been better, but the nature of the communication was very much between two processes and I didn't want to weaken that.

Coordination can be expensive and certainly, actors or some sort of channel-based model (which is my preference), have their place, but when you have light-weight communication and coordination primitives, it's not a given.

The ZeroMQ documentation suggests that you should use ZeroMQ to communicate between threads within the same process, as well as between processes. Given that context it wouldn't make sense for it to be thread safe. That would be too much overhead for no benefit.

ZeroMQ is very poor at communicating between threads within the same process, because messages are just binary data. The messaging library I use shouldn't dictate the way that I program; the dog wags the tail.

That would force me into an asynchronous style, but in my case, I was able to convey intent much better using threads.

This complaint is borderline non-sensical to me. Maybe Haskell is magic?

If ZeroMQ transported anything besides binary blobs, it would be marshaling objects or enforcing some kind of encoding to your data, which would dictate much more about your program. Are you upset that you had to use some Haskell.Vector to void* to Haskell.Vector code?

Are threads in Haskell not asynchronous by nature? Do you get synchronization for free somehow?

OK, in Haskell, data is immutable. If you want to pass data from one thread to another, there is no point in serializing it, sending it between threads and then deserializing it. You can just pass a reference. The data is also not typed, but that is potentially something you could fix in the bindings.

I know that in Erlang, you do pass data by exchanging blobs, but the difference is that Erlang has per Erlang-process heaps. Haskell has one heap, so there is no reason to do this.

It comes down to this, the way that you communicate between threads in the same process is different than the way that you communicated between processes.

> Are threads in Haskell not asynchronous by nature? Do you get synchronization for free somehow?

I'm not talking about what's happening underneath. I'm talking about the style that code is written in. I suppose, I could use a coroutine style monad over ZMQ's asynchronous API and create pseudo-threads, but they wouldn't get pre-empted, so you'd have to be careful to ensure that they didn't stave.

It depends on your service multiplexing scheme. If you are using something like NodeJS or EventMachine, then thread safety provides nothing but only increased complexity and slowness.

And 0mq is for inter-process communication. While you are running a server with many threads in one process, you usually don't need it.

I guess I'm arguing that this is a premature optimization.

While I'm arguing that modern libraries can avoid threads...

Since you have Haskell and Actors, but worrying thread-safety of some other libraries. You can consider a single-threaded/multi-process deploy.

By setting the thread pool to 1, Actors in a single process can still handle tens of thousands of concurrent connections, while there is surely no thread-safety problem.

As I've said elsewhere in this thread, the messaging library shouldn't dictate my programming style.

It seems to me that ZeroMQ is very good if you are creating a messaging network with load balancing, fanout, and fanin, however I don't think what most people are using HTTP services for matches to ZeroMQ full capabilities. That being said, it is very easy to get started with ZeroMQ. The whole idea of making everything messaging oriented seems to make sense especially if you think message passing is a good way to simplify complex systems that need high levels of concurrency. The issues with threading, forking, and contexts make some things tough with ZeroMQ, see: http://250bpm.com/blog:23. I stopped hacking on ZeroMQ when i was unable to trouble shoot some hard to reproduce crashes.

Thrift has been my backend RPC tool of choice both at work and on personal projects. It's been terrific for building composable backend services with tightly integrated language support.

Pros: - Explicit contracts through the IDL. I despise implicit contracts in service oriented code. - Code generation for many languages. - Server and client interfaces that don't require as much ceremony as HTTP. - Fast binary protocol

Cons: - Sometimes the generated code isn't exactly what you want (can feel a bit Java centric) - Binary protocol isn't human readable, can be harder to debug. - Stream oriented calls aren't very feasible without some sort of home-grown chunking solution. This is why tools like Cassandra tell you that the whole request must be able to be held in memory.

Thrift is definitely worth a look if you're shopping for an RPC tool: http://diwakergupta.github.io/thrift-missing-guide/

Can't you use framed transport to handle streaming better?

(my previous company gave up on thrift when we discovered it couldn't do recursive data structures (i.e. trees), and made our own interface -> service layer on top of protocol buffers instead).

As far as I know, you still have to read the entire response into memory for parsing. You could write your own streaming parser, but as far as I know the main compiled bindings don't support streaming response handling. I would love to be proven wrong though!

I really love ZeroMQ. We used it at LogNormal for all of the back end RPC. We did run into one issue though. If your 0MQ servers (the ones that call bind) run on a virtual host, as do your clients (the ones that call connect), and if that virtual host reboots, then the clients do not automatically reconnect to the server. I didn't quite have the time to investigate this, but from what I could tell, it only happens with virtual hosts across a reboot. It may have to do with iptables rules that we have in place (we have an IP whitelist), but I can't be certain of this.

I've never come across this. I'm able to kill my server (the one with the REP sockets), submit a form on the web page, have the HTTP request in the browser hang (that under the hood proxies to ZeroMQ), start up the server, and have the request in the browser get replied to almost immediately. But I've only been running ZeroMQ in topologies with two nodes at most, no advanced routing etc.

when you say kill the server, do you mean kill the server process or kill the box that the server runs on? I have no problem with the former, but I have a problem with the latter, and I can't tell why. It's pretty consistent.

It's a TCP issue. When the other end of the connection goes away silently (power goes of, cable is cut off etc.) TCP doesn't inform ZeroMQ that the connection is broken, so ZeroMQ uses it furthe without trying to handle the error.

It's only an issue with TCP if you forget to enable keep-alive (http://www.gnugk.org/keepalive.html). Keep-alive solves this problem, and it solves other problems too, like when your network equipment decides a connection is no longer needed and disables it. Or, you can implement application-level heartbeats--whenever you send you have the chance to recognize that your peer has gone away.

That's an interesting thought. I need to look into whether I have TCP keepalive turned on for my private interfaces.

I'd call that a characteristic of TCP, not an 'issue'.

I mean "kill -9" on the server process. I'm not under that much load, though. From reading the other comments, my suspect is REQ/REP socket lockstep, which I've never encountered, but that I guess would happen if something dies while the REQ server waits for a reply that it never gets. Do you have timeout handling in your system?

I have no problem with kill -9. I use PUB/SUB sockets. I only have problems with reboots.

I suspect it is something in you setup -- that is a use case that ZMQ is specifically very good at... the fact that you can start the client before or after you start the server (doesn't apply to inproc) is awesome.

yeah, starting/stopping the server works without a problem. rebooting the vm on which the server runs is what causes the problem.

Have you compared the traffic using Wireshark or tcpdump?

unfortunately I can't do that while the vm is booting up, and after bootup there is no traffic, so nothing to look at.

Without knowing details this sounds very much like the stuck clients issue Armin Ronacher complains about here: http://lucumr.pocoo.org/2012/6/26/disconnects-are-good-for-y... (discussed here https://news.ycombinator.com/item?id=4161073)

In a nutshell if the server goes away while a client is in recv the client will never know and waits forever.

REQ and REP have an evil, evil failure mode if they get out of sync, as I recall. The lockstep of request and reply is rather easy to break.

At JacobsParts we experienced this. Every few weeks our central API would just stop responding. We were using a threaded python zeromq server and the REP "sockets" would get stuck, even after the client was gone. Enough stuck sockets and all threads were hung. Now we use python gevent for an asynchronous server and we use ROUTER, DEALER instead of REP, REQ. Actually the clients can use either DEALER or REQ depending on if they are asynchronous or threaded. Anyway the results have been fantastic and stable for a year now. When there are many small requests the performance benefit of ZeroMQ is just amazing. Can hardly tell the difference between a remote call and a local function call!

+1 on sticking to DEALER/ROUTER on top of a green-thread system like gevent. This is how zerorpc works [1], and it is the backbone of the dotCloud platform.

Also note that most of the time PUB/SUB and PUSH/PULL are not a good idea either. The same results can be usually be achieved by returning a stream on top of ROUTER/DEALER (this is what zerorpc does). The performance gains of custom topologies are great in theory, but in a typical modern web or mobile stack, they are not worth the extra effort and lack of flexibility. The single best change we made to dotCloud's architecture was move away from custom topologies and stick to DEALER/ROUTER.

[1] http://github.com/dotcloud/zerorpc-python

I'm curious, did you have some sort of timeout mechanic for your REP sockets? I don't, but I've never had problems either because I don't have that much traffic.

AFAIK ZeroMQ doesn't support any kind of timeout on REP sockets. It could be hacked in with signals or a watchdog thread in a multiprocessing setup, but that's ugly. If you're going to all that trouble it seems much cleaner to just move to DEALER/ROUTER.

I do like the way REQ/REP keeps track of the requests and has RPC semantics, though. Instead of blocking, you could always poll and time out instead.

The "lazy pirate pattern" was added to the Guide a while back. It explains how to deal with this robustly.


Are you referring to the need to use a timeout? http://lucumr.pocoo.org/2012/6/26/disconnects-are-good-for-y...

Anyone used a ESB(biztalk, websphere) rather than a pure messaging solution. What's the difference?

Expensive consultants, low-quality error reporting, and a bunch of useless tools that make management feel like they can program (and then leave you to fix their crap)?

I presume there must be some upside, but I never saw it.

ZeroMQ is supposed to be a superior alternative to a broker with reduced hops and avoids a single point of failure.

More details here: http://www.zeromq.org/whitepapers:brokerless

I've been using Avro because of its built-in support for Authentication via SASL. Does anyone know what the author is referring to when he says work is in progress for providing encrypted transports in ZeroMQ? Is there an issue that I can follow somewhere for this?

CurveZMQ - authentication and encryption library


CurveZMQ was recently announced by Pieter Hintjens (one of the 0mq contributors) on his blog:


HTTP does have pipelining since the 90s.

I did look up HTTP pipelining before rolling with this, it is severly limited. It can only really do sequential requests concurrencly, not an arbitrary mishmash of requests and replies as ZeroMQ can. Also, pipelining should only do idempotent requests, apparently.

My only source for this is the wikipedia article [1] though, do you have some additional info about this?

[1] http://en.wikipedia.org/wiki/HTTP_pipelining

Does anyone have any experience using ZeroMQ with larger messages? For instance if you need to send messages that are several megabytes for rpc style calls.

I just remember seeing anecdotal comments about it not working well, but never really dug into whether that was sound.

I've used it to stream an uncompressed video feed (essentially a special purpose webcam) over LAN as a quick and dirty solution for a research project. It worked out ok after I set a high water mark and forced it to drop data when it couldn't keep up. Before I figured that out, messages tended to pile up and crash my process by exhausting memory.

Can't say I'd recommend it, but it can be done, depending on the needs of your project.

Why not use a secondary storage for the large chunks of data, and pass a unique identifier to zeromq?

I don't follow the reasoning; wouldn't it be more sensible to compare ZeroMQ vs. TCP instead of ZeroMQ vs. HTTP? If you can so easily replace HTTP, why the fuck did you choose it in the first place instead of a plain TCP protocol?

ZeroMQ is not really at TCP level, ZeroMQ is about sending messages to and from various semantic endpoints, such as REP and REQ sockets, specifically made for doing RPC. If I was using TCP, I'd have to do a _lot_ of work myself (keeping the connection up, throttling locally when it's not up, keeping track of which messages goes where, etc).

ZeroMQ is an impressive project and I like reading Martin's (passionate!) blog posts. Haven't really needed it though. I hope I'll get a chance to use it to build a distributed high-performance system one day.

On a related note, it looks like Crossroads I/O is dead.

Crossroads I/O has been superseded by nanomsg: https://github.com/250bpm/nanomsg

It's still pre-alpha but has a ZeroMQ compat layer already.

I see that nanomsg is still C as opposed to C++.


Why dead, why superseded?

IIRC: Crossroads I/O was forked from ZeroMQ by its original author, who subsequently decided to start from a clean slate with nanomsg. His blog has some really good insights on various design decisions and "lessons learned" from ZeroMQ: http://250bpm.com/blog

HTTP works fine for me. Seems like a nice article though.

How do you manage connections? Do you use keep-alive or not? If so, you have some sort of connection pool mechanism in order to be able to perform requests in parallel? These two problems is the main reason for why I switched to a MQ instead of using HTTP.

(Hopefully I managed to not make this seem like flaming, I'm genuinely curious.)

I think it was meant that switching to 0MQ instead of HTTP provides little benefit at this stage for that person. The great thing about HTTP is that you can quickly whip up services that will work without worrying about performance until later. I'm sure the same thing can be said about 0MQ once you are fluent in it, but it is not ubiquitous as HTTP.

In other words, it can be considered a premature optimization

Note that HTTP pipelining will let you have multiple outstanding HTTP requests on the same TCP connections, although it comes with some caveats: http://en.wikipedia.org/wiki/HTTP_pipelining

@augustl: We manage connections with curl and requests(python). We haven't seen any bottlenecks here yet, so we haven't thought to revisit it.

We do, however, use ZeroMQ for other stuff. But I'll always reach for HTTP for an API, unless there's a need to do otherwise. It's super simple to get running, interface and debug with. Every backend engineer knows how to use curl.

Have you considered/tried using a decent HTTP client + HAProxy?

I have, but the fundamental limitation of sequential req/rep on a connection with HTTP pushed me to find alternatives.

Do you have any recommendations for decent HTTP clients?

This. 0MQ is an optimisation and as everyone knows, these optimisations should be delayed until the last minute.

We're still shifting services over HTTP off a mix of Apache and IIS boxes. We can do 80 million service requests a day without hitting 5% of our capacity with commodity hardware.

These service requests are extremely complicated calculation and scoring algorithms as well so not some half-arsed CRUD API either.

We just do it in C# and C++ rather than Python/Ruby etc as that's where we gain the performance advantage.

What about interoperability with external services? With HTTP, you could just open your APIs to the world, when it makes sense.

OP here. In the blog post, I use a traditional HTTP routing library on the server, and I return HTTP-like responses. Since I didn't reinvent that part, it's actually pretty trivial to either use the same code and package it into a war file, or just fire up a Jetty in-process if you only have one server instance.

Any experience with json-rpc? I'm trying to decide between the two.

I'm acutally doing json-rpc over ZeroMQ using a scheme similar to the one described above to do file conversion tasks.

A ZeroMQ proxy using a ROUTER/DEALER pair with a bunch of REP sockets in the background.

The clients use a simple REQ socket.

All in plain old C using ZeroMQ and jansson for JSON while conforming to the json-rpc 2.0 spec.

If I had to do that, I'd probably go with a queue on Redis or Postgres. "Cogs bad".

What are the advantadges of ZeroMQ when you're still on one computer?

That was very helpful thank you.

I see a lot of bare IP addresses and whatnot in ZeroMQ examples. Is it easy to integrate with a naming system like Zookeeper?

The ZeroMQ guide has a number of sections on discovery: http://zguide.zeromq.org/page:all#Discovery However, you could also use something like zookeeper.

The guide is a fascinating read. Even if you never write anything using ZeroMQ, it's a very useful intro to designing concurrent message passing systems.

What I got from reading that is the author has no idea how to solve the problem of many frontends and many backends of many services on a large scale (same feeling I got from reading the section about heartbeating.)

Interesting. Can you enlighten us with more detail?

The author lays out a dozen different, terrible ways to discover services, only one of which is even remotely suitable. I associate that kind of stream-of-consciousness documentation with people who are making it up as they go along.

Anyway, the proposed solution of a full-mesh FE-to-BE heartbeat network over UDP, switching to TCP in case of idleness, is not going to scale. That just guarantees that you are going to run out of file descriptors in case of a packet loss event, as everyone upgrades their heartbeat protocol to TCP.

What, in your opinion, is the correct way to implement heartbeats in such a message oriented system?

http is a bad protocol for internal services due to arbitrary length text headers with big parsing overhead and lack of multiplexing of packets per connection by default. Even FastCGI is much better choice.

Erlang's messaging protocol, Protocol Buffers, Msgpack and other carefully serializing solutions trying to reduce parsing and communication overheads.

ZeroMQ seems like natural choice for the core of so-called messaging middleware, because, well, it's what it was designed for.)

And, of course, avoiding JVM leads to great reduction of resource wasting.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact