IRC Log


Thursday July 14, 2011

[Time] NameMessage
[03:00] cremes rev: inproc works great on windows; i think you probably mean ipc
[03:00] cremes that isn't supported on windows yet
[03:03] cremes pauldix: you need to reread the section on message envelopes: http://zguide.zeromq.org/page:all#Request-Reply-Envelopes
[03:04] cremes in order for your example to work, you need to pass the "null" message part that separates the routing envelope from
[03:04] cremes the message body
[03:04] cremes i'll put a small code example in a comment on your gist
[05:35] sustrik mikko: pong
[05:35] sustrik pieter_hintjens: pong
[06:38] pieter_hintjens sustrik: :-) good morning, I already backported the 228 fix to 2.1/2.2
[06:39] sustrik great, thanks
[09:41] mikko sustrik: was wondering 3.x
[09:41] mikko is the protocol from 3.x future compatible?
[09:42] mikko 2.x -> 3.x makes a very painful migration due to protocol not being supported
[09:49] sustrik mikko: hi
[09:49] sustrik hard to say
[09:49] sustrik the protocols can change, eg. if acks are introduced to PUSH/PULL
[09:50] sustrik these have to be passed on the wire
[09:50] mikko 2.x -> 3.x jump seems pretty annoying from user's perspective
[09:50] sustrik it is
[09:50] mikko needs downtime in many scenarios
[09:51] sustrik no idea what else to do though
[09:51] sustrik any ideas?
[09:52] mikko is it possible to include one byte for protocol version?
[09:52] mikko and handle that accordingly
[09:52] sustrik how would that work?
[09:52] mikko maybe a bit late now when 3.x is out
[09:53] pieter_hintjens sustrik: I've been thinking about this
[09:53] sustrik refusing the connections with incompatible version numbers>
[09:53] sustrik ?
[09:53] pieter_hintjens we already discussed how to do it
[09:53] mikko it would probably need a bit more design in handling protocol in backwards compatible way
[09:53] eintr 2.x and 3.x are frame / wire compatible though, right?
[09:53] pieter_hintjens mikko: it's not too late, this can be added at any point
[09:53] sustrik eintr: nope
[09:53] pieter_hintjens eintr: nope
[09:53] sustrik :)
[09:53] eintr okie then, ignore me (haven't had time to actually check out 3.x yet) :)
[09:53] pieter_hintjens mikko: the difficulty with designing the 3.0 protocol is it's not documented
[09:54] mikko we are quickly going to alienate users if every major version requires shutting down operation
[09:54] pieter_hintjens mikko: obviously
[09:54] sustrik so, what's the propsal?
[09:54] pieter_hintjens well, first off, document the proposed changes to 3.0 already
[09:54] pieter_hintjens so we know what we're actually talking about
[09:55] pieter_hintjens second, add version detection to the protocol
[09:55] pieter_hintjens so that a 3.0 stack can detect a 2.1 peer safely
[09:55] pieter_hintjens and then that a 2.1 peer can safely detect / disconnect a 3.0 peer
[09:55] sustrik how does that solve the problem?
[09:56] pieter_hintjens those are necessary first steps
[09:56] sustrik ok, what next?
[09:56] pieter_hintjens then, in 3.0 we can look at supporting the old protocol syntax
[09:56] pieter_hintjens since the semantics haven't changed
[09:56] sustrik ugh
[09:56] pieter_hintjens ugh, yes
[09:56] pieter_hintjens or, we can make bridges
[09:56] pieter_hintjens but neither of these are even possible today
[09:57] pieter_hintjens we can't even safely detect version incompatibility
[09:57] sustrik that can be done, but the main question is: who's going to maintain 2.0 protocol in 3.0
[09:58] pieter_hintjens it shouldn't require *maintenance*
[09:58] sustrik ?
[09:58] pieter_hintjens if you start on the basis of "I don't want to alienate all my users" then you take a sensible path
[09:59] pieter_hintjens you change the protocol format but not the semantics
[09:59] pieter_hintjens that means it's simple to support both old and new
[09:59] sustrik well, the semantics do change
[09:59] pieter_hintjens then when you have migrated the bulk of people to the new protocol version
[09:59] pieter_hintjens then you start to add new semantics
[09:59] sustrik what else would be the point of changing the protocol
[10:00] pieter_hintjens e.g. label flag, and different ID formats are not semantic changes
[10:00] pieter_hintjens adding version header is not a semantic change
[10:00] sustrik ok, true
[10:00] pieter_hintjens req-rep reliability is a semantic change but layered on top in any case
[10:00] pieter_hintjens so it's doable, if you actually decide it's worth doing
[10:01] sustrik i have no resources to do that
[10:01] sustrik same problem as with explicit identities
[10:01] sustrik the worth of doing so exceeds my bandwidth
[10:01] pieter_hintjens you have resources to run but not walk?
[10:01] pieter_hintjens e.g. spending 10 minutes to properly document requirements for 3.0 protocol
[10:01] pieter_hintjens so that other people can help design a better WLP
[10:02] sustrik combining old and new functionality means supporting a lot of complex infrastructure
[10:02] pieter_hintjens lack of versioning is a major weakness, due simply to "I don't care about this"
[10:02] pieter_hintjens combining old and new means taking more time to move, that's all
[10:02] pieter_hintjens like I said, you do a syntax change in one step, semantic changes in later steps
[10:02] pieter_hintjens not at the same time
[10:03] sustrik how does that solve the problem?
[10:04] pieter_hintjens it lets people upgrade safely to 3.0 and then onwards
[10:04] pieter_hintjens if you make a barrier to upgrading people will stay on 2.1
[10:04] pieter_hintjens being unable to mix old and new is a major barrier
[10:04] pieter_hintjens it means larger architectures have to do a big bang upgrade
[10:04] pieter_hintjens you're forcing people to test the new functionality all in one go
[10:05] sustrik still better then asking them to test new incompatible version several times
[10:05] pieter_hintjens ?
[10:05] pieter_hintjens what you want is to deliver new functionality (e.g API, subscription forwarding, req/rep reliability) on the OLD protocol so people can test it safely
[10:06] sustrik you can't
[10:06] sustrik new fuctionality needs new protocol
[10:06] sustrik like subscriptions on the wire
[10:06] pieter_hintjens nope, it doesn't
[10:06] pieter_hintjens those are socket pattern semantics
[10:06] pieter_hintjens old code will work unchanged
[10:06] sustrik it won't
[10:06] sustrik subsctriptions will accumulate in the upstream node
[10:06] pieter_hintjens xsub/xpub affects my push/pull sockets?
[10:07] sustrik and finally it will run out of memory or something
[10:07] pieter_hintjens well, this is also a design choice
[10:07] sustrik new pub/sub doesn't work with old pub/sub
[10:07] pieter_hintjens you _could_ make it compatible
[10:07] sustrik to do that i would have to change 2.0 to speak 3.0 protocl
[10:07] pieter_hintjens e.g. totally separate socket types
[10:07] sustrik ZMQ_PUB1 vs. ZMQ_PUB2
[10:07] pieter_hintjens sure
[10:07] pieter_hintjens why not?
[10:08] sustrik who's going to maintain it?
[10:08] rimas hi guys, could anyone help me out with actionscript TCP wrapper for zmq?
[10:08] rimas basically I'm having trouble sending data to zmq
[10:08] pieter_hintjens rimas: you might have to track down the author...
[10:08] pieter_hintjens but I know it works, have seen it in action
[10:09] rimas well, recv does work fine, however send doesn't :(
[10:09] pieter_hintjens sustrik: it comes down to making smaller changes, getting them live faster, and not making major breaks
[10:09] pieter_hintjens and not making multiple major breaks at once
[10:10] rimas maybe you guys can explain a bit better on how the whole handshake works with zmq?
[10:10] pieter_hintjens rimas: what kinds of sockets are you using?
[10:10] sustrik i have no problem with creating new patterns and handing them out to people who are willing to maintain them
[10:10] pieter_hintjens sustrik: "maintenance" is a kind of fake problem afaics
[10:10] sustrik so, if anyone wants ZMQ_OLD_PUN & ZMQ_OLD_SUB, no problem
[10:10] rimas well, i've got python zmq.REP
[10:10] pieter_hintjens code that is used, gets maintained
[10:10] rimas as a server
[10:11] sustrik rimas: there's no handshake, each party just sends one empty message after connecting
[10:12] pieter_hintjens rimas: I don't think anyone here knows the AS binding. I'd suggest making a test case, post to the email list, perhaps contact the author
[10:13] rimas ok, thanks guys
[10:13] pieter_hintjens sustrik: essentially, with 3.0 as it exists today, existing 0MQ users cannot easily upgrade, only new ones can use it
[10:13] sustrik yes
[10:13] pieter_hintjens however, until people are using 3.0 it will not become stable
[10:13] pieter_hintjens and until it is stable, new users will not take the risk of using it
[10:13] sustrik sure
[10:14] sustrik what does that have to do with the fact i can't handle all the maintenance
[10:14] pieter_hintjens conclusion: unless you solve the problem of upgrading, none of the new functionality in 3.0 will ever get widely used
[10:14] pieter_hintjens and if it's not used, no-one will maintain it
[10:14] pieter_hintjens are you doing a lot of maintenance on 2.1 now?
[10:14] sustrik nope
[10:15] pieter_hintjens well, why do you keep raising this as an issue then?
[10:15] sustrik i don't care about 2-1
[10:15] pieter_hintjens I think I've proven that maintenance is a solved problem
[10:15] sustrik actually, i don't care about 3-0, you've agreed to maintain it
[10:15] pieter_hintjens indeed
[10:16] pieter_hintjens but that doesn't mean it'll get used
[10:17] sustrik i guess the people who want to use 2.1 protocls and 3.0 functionality should produce the patches
[10:17] pieter_hintjens if you want people to use the code you're writing, you can't just push it and then say "I'm not maintaining it"
[10:17] pieter_hintjens that won't work
[10:17] pieter_hintjens you have made it pretty impossible for anyone to contribute to the protocol
[10:17] pieter_hintjens so no-one will propose changes to that
[10:17] pieter_hintjens I don't want to be negative
[10:18] pieter_hintjens 3.0 is pretty cool, I like the simplifications
[10:18] sustrik i personally believe that 3.0 should have been dropped in the beginning
[10:18] sustrik i've proposed to do so at the time
[10:18] sustrik and everyone went like "don't drop it"
[10:18] sustrik so here it is
[10:18] pieter_hintjens what was the alternative?
[10:19] sustrik stabilise 2.x
[10:19] pieter_hintjens no future development of 0MQ?
[10:19] sustrik exactly
[10:19] pieter_hintjens 2.x is stable
[10:19] pieter_hintjens was stable already
[10:19] sustrik support existing users
[10:19] pieter_hintjens all that's happening
[10:19] sustrik why do we need 3.0 at all?
[10:19] pieter_hintjens you enjoyed making it, I assume
[10:19] sustrik i can do that work as a private experiment
[10:19] pieter_hintjens otherwise you'd have done stuff like make 0MQ robust for Internet use
[10:20] sustrik yes
[10:20] pieter_hintjens we still have 0MQ/2.2 as an option for thart
[10:20] pieter_hintjens we can make 2.2 interoperate with 3.0
[10:20] sustrik right
[10:20] pieter_hintjens maybe 3.0 is simply too far ahead
[10:20] sustrik i don't see much point in imposing new functionality on users that don't want it
[10:20] sustrik yes, something like that
[10:21] pieter_hintjens well, people do want the subscription forwarding
[10:21] pieter_hintjens and you need running code for the SP framework
[10:21] pieter_hintjens the reasons for making 3.0 were / are sound afaics
[10:22] sustrik but it doesn't play well with stability of the product
[10:22] sustrik it's stable vs. experimental trade-off
[10:22] pieter_hintjens this is normal, and fine
[10:22] pieter_hintjens we have our product versions and cycles well defined and they work
[10:23] pieter_hintjens the only flaw here is the difficulty of mixing 3.0 code with 2.x code in one architecture
[10:23] pieter_hintjens that's all
[10:23] pieter_hintjens I can't test any of the new stuff without taking it all
[10:23] pieter_hintjens that's too costly and risky
[10:23] sustrik yes
[10:23] pieter_hintjens I'd say, either we get 3.0 to speak nice to 2.1, or we do this in 2.2
[10:23] pieter_hintjens one or the other
[10:23] pieter_hintjens that ensures 3.x will get used
[10:24] sustrik i would vote for 2.2
[10:24] pieter_hintjens that makes sense
[10:24] pieter_hintjens it means the 3.x codebase doesn't get messier
[10:24] pieter_hintjens and "speaking with" can be strictly constrained
[10:25] sustrik yes
[10:25] pieter_hintjens 2.2 is in good shape, it has all the 2.1 patches plus a couple of other changes
[10:25] pieter_hintjens let's think about this for a while, there's no real hurry
[10:26] sustrik ok
[10:26] pieter_hintjens if you could send me some WLP-patches for 3.0 we can think about how to add version detection to the protocol in a safe way
[10:27] sustrik i can send you the LABEL patch
[10:28] sustrik the sub forwarding patch is actually a big bunch of patches that toueches almost every file in the codebase
[10:28] pieter_hintjens i'm only concerned with ZMTP at this point
[10:28] pieter_hintjens we should also expand that to document the patterns
[10:28] sustrik ok, let me check
[10:29] pieter_hintjens mikko: does the idea of using 2.2 as a bridge version make sense?
[10:31] sustrik LABALE patch: ab99975ad44ed0fe9ab6
[10:31] sustrik LABEL
[10:32] pieter_hintjens sustrik: at http://rfc.zeromq.org/spec:13#toc16
[10:33] pieter_hintjens what changes does the label function make to that?
[10:34] sustrik drop more & final
[10:34] sustrik introduce 'flags' instad
[10:34] sustrik flags = 0x00
[10:34] sustrik flags = 0x01
[10:34] sustrik flags = 0x80
[10:34] sustrik flags = 0x81
[10:35] sustrik the lowest bit being LABEL flag
[10:35] pieter_hintjens we still use flags for more/final, right?
[10:35] sustrik the highest bit being MORE flag
[10:35] sustrik yes
[10:35] sustrik however, it is split
[10:35] pieter_hintjens more-label, final-label, more-data, final-data?
[10:36] sustrik othre way round
[10:36] sustrik 0x80 = more
[10:36] sustrik 0x01 = label
[10:36] pieter_hintjens yeah, I mean semantically
[10:36] sustrik yes
[10:36] pieter_hintjens in terms of what a message is
[10:37] sustrik however, the two are mutually exclusive
[10:37] pieter_hintjens 0..n label frames, 0.n data frames
[10:37] sustrik so more-label doesn't happen
[10:37] sustrik exactly
[10:37] pieter_hintjens so label always implies more?
[10:37] sustrik yes
[10:37] pieter_hintjens can you have a message consisting only of label frames?
[10:37] sustrik maybe it's wrong
[10:38] pieter_hintjens this is why you want protocols discussed independently of code...
[10:38] pieter_hintjens does 0MQ allow an empty message as in zero frames?
[10:38] pieter_hintjens I don't think it does, does it
[10:38] sustrik yes
[10:38] pieter_hintjens that's why ZMTP says *more-frame final-frame
[10:39] pieter_hintjens so it'd be *label-frame *more-frame final-frame
[10:39] sustrik yes
[10:39] pieter_hintjens ok
[10:39] pieter_hintjens if you read the next section of that page, it proposes a mechanism for version detection
[10:41] sustrik it's not compatible with 2.x protocol
[10:42] pieter_hintjens it's not compatible with 2.1 code
[10:42] pieter_hintjens which would break
[10:42] sustrik yes
[10:42] pieter_hintjens but that can be patched trivially, IMO
[10:42] pieter_hintjens i.e. reject such connections as invalid
[10:42] sustrik the problem is that you are back to the original problem
[10:42] sustrik you have to deploy 2.2 in one go
[10:42] sustrik on gradual upgrading
[10:42] pieter_hintjens so, my proposal is that 2.2 actually do version detection
[10:43] pieter_hintjens i'm willing to make this, as far as I can
[10:43] pieter_hintjens anyhow, design first... 2.2 should be able to speak to 2.1 apps
[10:43] sustrik yeah
[10:43] pieter_hintjens and 2.2 can also speak to 2.2 apps, using zmtp/1.1
[10:44] pieter_hintjens and 2.2 can then speak to 3.0, using zmtp/1.1
[10:44] sustrik 1.2?
[10:44] pieter_hintjens no need, 1.1 will have the layering for xpub/xsub, etc. on top
[10:45] sustrik ok
[10:45] pieter_hintjens I assume we can either add this explicitly with socket pattern validation, or stick with "if you mix the wrong stuff it breaks, too bad" approach
[10:45] pieter_hintjens all that really matters is you can upgrade parts of your network piece by piece
[10:46] pieter_hintjens this means the upgrade path is (a) move to 2.2, which is relatively safe because it's the same stable codebase for 95%
[10:46] mikko pieter_hintjens: i think ideally from user's perspective tehy shouldnt know about versions
[10:46] pieter_hintjens and (b) then start moving to 3.0 piece by piece
[10:46] pieterh mikko: the technical user is going to be very aware of versions because they bring specific functionality
[10:46] pieterh and risks
[10:47] pieterh no avoiding that - new code is always riskier than old code
[10:47] pieterh mikko: I made a lovely hard hack yesterday
[10:47] pieterh nothing to do with 0MQ...
[10:48] pieterh while true; do if test -d /media/*/Android; then banshee --play; else banshee --pause; fi; sleep 1; done
[10:48] pieterh dock my mobile phone, music starts playing... undock it, music stops immediately
[10:48] jsimmons lol
[11:01] mikko pieterh: heheh
[11:02] pieterh mikko: I first thought, "buy a kinnect and use it to see if I'm at my desk or not"
[11:02] mikko pieterh: it would be fairly simple to do with openni
[11:02] mikko gesture based music control
[11:02] pieterh then, "hook into screensaver and use that to stop/start music"
[11:03] pieterh nice thing is when someone calls me, I just pick up the phone and answer, and the music magically pauses...
[11:30] pieterh sustrik: wrt label frames
[11:30] pieterh can these be used for all socket patterns?
[11:35] sustrik pieterh: yes
[11:36] sustrik it's a generic mechanism for tagging messages
[11:36] pieterh ok
[11:36] sustrik i borrowed the concept from MPLS
[11:37] pieterh so version numbering...
[11:37] pieterh if we want to do socket pattern validation we also need to send some kind of greeting message
[11:38] pieterh hmm, we could already disable identities in this wlp
[11:39] sustrik you'll break 2.x compatibility that way
[11:39] sustrik which would destroy the whole point
[11:39] pieterh well... :-)
[11:39] pieterh not really
[11:39] pieterh imagine a 2.2 broker taking to 2.1 and 3.0 clients
[11:40] pieterh the 2.1 clients can still use explicit identities, 2.2 will support that
[11:40] pieterh 2.2 to 2.2 won't, and 2.2 to 3.0 won't either
[11:40] sustrik 3.0 *does* support explicit identities
[11:40] pieterh ok
[11:41] pieterh so then we have to send a more complex greeting
[11:41] pieterh i'll propose something, we can beat that into shape
[11:41] sustrik ok
[11:59] pieterh sustrik: is there any value in adding a type octet to label frames?
[11:59] pieterh right now the syntax of these is undefined
[12:04] sustrik pieterh: i would say the actual semntics of the labels is up to the messaging pattern
[12:04] pieterh ok... makes sense
[12:04] pieterh I've pushed a draft then
[12:04] pieterh http://rfc.zeromq.org/spec:15
[12:04] sustrik where?
[12:04] sustrik let me have a look
[12:04] pieterh its simpler than spec:13
[12:06] sustrik hm, it's not compatible with 2.1
[12:06] pieterh :-) of course not ...
[12:06] sustrik so how would the upgrade process work?
[12:07] pieterh ah, ok, so this spec lets us make a 2.2 that speaks both old and new protocols
[12:07] pieterh since the semantics are the same, just wire format changes a little
[12:07] pieterh I'll need to check that 2.1 correctly kills peers that send invalid identities
[12:08] pieterh it requires that the 2.1 peer talks to the 2.2 peer first, not the other way around
[12:08] sustrik what's an invalid identity?
[12:08] sustrik multi-part one?
[12:08] pieterh one starting with a null byte, for zmtp/1.0
[12:08] pieterh or, yes, a multipart one
[12:09] pieterh both are invalid according to zmtp/1.0
[12:09] sustrik ok, i see
[12:12] pieterh in 2.1, a subscriber sends a greeting right away, doesn't it
[12:12] sustrik it sends the identity right away
[12:13] pieterh ok, so I also added a sender socket type in the greeting
[12:14] sustrik ok
[12:15] pieterh this'll do for now, I should be able to make the changes to 2.2 if no-one else feels like it
[12:15] pieterh there'll be some cases where mixing versions does not work
[12:50] guido_g pieterh: of by one error in spec detected! "Bits 1-6: Reserved. Bits 1-7 are reserved for future use and MUST be zero."
[12:50] guido_g hi btw
[12:51] guido_g what happens when length is 0?
[12:54] sustrik guid_g: invalid message
[12:55] guido_g ok
[13:49] pieterh sustrik: it's one thing I'd like to change, make frame size exclude any frame meta data (i.e. flags)
[13:49] pieterh so that frame size 0 is valid
[13:49] sustrik why so?
[13:49] pieterh the current design is consistently confusing
[13:50] sustrik there are two layers in it actuallyu
[13:50] pieterh it confused me, and I've seen several people hit it exactly the same way
[13:50] sustrik there's message delimitation protocol (the size + data)
[13:50] sustrik and labeling protocol (flags+data)
[13:50] sustrik the delimitation protocol doesn't know about labeling protocol running on top of it
[13:50] pieterh well, that is an explanation but it's not helpful
[13:51] pieterh consistently, people ask "why is a size zero not valid?"
[13:51] sustrik it is valid frame
[13:51] sustrik not a valid messafe part
[13:51] sustrik message
[13:51] pieterh we don't actually have two protocols
[13:51] sustrik you can split the doc if you want
[13:51] pieterh and no other protocols do this
[13:51] pieterh that would be insane
[13:52] pieterh take web sockets as an example
[13:52] pieterh frame header + frame data
[13:52] pieterh header includes data length + meta data
[13:52] pieterh that is what people expect, it's sane
[13:52] pieterh to define two protocols just to define a frame is insane
[13:52] pieterh and so, no-one expects it, so it's consistently confusing
[13:52] pieterh that's my experience
[13:52] sustrik easy to process in hardware
[13:53] pieterh i understand that is the underlying rationale but...
[13:53] pieterh adding 1 or 2 doesn't change anything IMO
[13:53] sustrik it binds the framing layer tightly with layering layer
[13:53] pieterh and the problem with that is?
[13:54] pieterh do you have any use case for reusing a 1-line framing layer?
[13:54] pieterh for various reasons but mainly future IEFT compatibility I'd really try to make it look more like the websockets header
[13:54] pieterh *IETF
[13:54] sustrik low level hardware
[13:54] pieterh unproven
[13:55] sustrik say a simple forwarder
[13:55] sustrik no need to know the labeling protocol
[13:55] pieterh philosophical arguments of "efficiency" vs. practical experience of "this design is confusing"
[13:55] sustrik come on, it's just a number
[13:55] pieterh and this is not SP, it's just a WLP for 0MQ
[13:55] pieterh it's a consistent cause of confusion
[13:55] sustrik feel free to change it in 2.2 then
[13:56] pieterh making a specific protocol for 2.2 would be kind of stupid
[13:56] sustrik you can change it in 3.0
[13:56] sustrik you are the maintainer
[13:56] pieterh sigh
[13:56] sustrik it's your decision
[13:56] pieterh sustrik: please don't waste my time
[13:57] pieterh really, if I'm explaining why the size field is confusing, accept my explanations and then think about it
[13:57] pieterh do *not* give me pointless arguments ending in "well, you maintain it, do what you like"
[13:57] sustrik i prefer to design based on technical arguments
[13:57] pieterh if the practical experience is outweighed by the philosophy, fine
[13:58] pieterh but do not just ignore feedback from your users
[13:58] sustrik in this case it's technical argument vs. make it look like websockets
[13:58] pieterh and do not, fgs, just tell them to patch your software
[13:58] pieterh I am not going to make changes to 3.0, you should understand that
[13:58] sustrik what do you want me to do then?
[13:59] sustrik imo the "make it look like websockets" argument doesn't make much sense
[13:59] pieterh I don't want you to do anything at all except listen to your users and use our accumulated experience to make 0MQ a better product
[13:59] pieterh that was not my main argument, it is incidental
[14:00] pieterh the websockets framing solves very much the same problem and comes from 1 year of argument on the HyBi list
[14:00] sustrik ok, so what the technical argument?
[14:00] pieterh the current design is consistently confusing
[14:00] pieterh sigh
[14:00] pieterh confusing designs are by definition bad
[14:01] pieterh layering the framing into two protocols is significant over-engineering that has a negative cost-benefit outcome
[14:01] sustrik say IP uses the same design
[14:01] pieterh not relevant
[14:01] sustrik IP is more common than websockets
[14:01] pieterh not relevant, no-one is making IP stacks today
[14:02] pieterh look, just tell me this is to allow super fast hardward parsing of frames
[14:02] pieterh discussion over
[14:02] pieterh *hardware
[14:04] pieterh if you turn protocol discussions into "well, you can change it in version X of the software, not my problem"
[14:04] pieterh you are effectively saying, "I'm not interested in interoperability"
[14:05] pieterh if you find yourself doing this to a significant community of users, your credibility as protocol designer is pretty dead
[14:06] reuben what's the downside of doing what pieterh is suggesting?
[14:06] reuben I mean, even if there -
[14:06] reuben >_>
[14:16] sustrik re
[14:16] sustrik sorry, got disconnected
[14:17] sustrik <sustrik> the problem behind the design is framing in transport-specific while labeling is not
[14:17] sustrik <sustrik> ie. sctp has framing of its own
[14:17] sustrik <sustrik> so it only needs the labeling part
[14:18] sustrik btw, there's one more thing i would like to do
[14:19] sustrik it has to do with building new patterns on top of 0mq
[14:19] sustrik currently, ROUTER can be used for that
[14:19] sustrik but it is basically hijacked XREP
[14:20] sustrik so i though of creating a socket type that would fit the role better
[14:20] pieterh for ROUTER?
[14:20] sustrik like, for example, it could send connection/disconnection messages etc.
[14:20] pieterh sure, there are a lot of potentially useful improvements to ROUTER
[14:21] sustrik once it is not bound to req/rep
[14:21] sustrik we are free to change it
[14:21] sustrik so that is suits the purpose better
[14:21] pieterh note that for ROUTER to make any sense it has to talk to other socket types
[14:21] sustrik my problem with that is that it breaks the semantics of existing patterns
[14:22] sustrik however, if the goal is to create new patterns
[14:22] pieterh well, what's the alternative to that?
[14:22] sustrik to problem disappears
[14:22] sustrik the semantics are up to the user space implementation of the pattern
[14:22] pieterh you need some other socket type (non-ROUTER) in 95% of patterns
[14:22] pieterh DEALER is most common
[14:22] pieterh so you could create SERVER and CLIENT
[14:23] sustrik you can create dealer-like semantics using generalised router socket
[14:23] sustrik no?
[14:23] pieterh where SERVER is ROUTER with extensions and CLIENT is like PAIR
[14:23] pieterh no
[14:23] pieterh you cannot use ROUTER as DEALER
[14:23] sustrik why so?
[14:23] pieterh srsly?
[14:23] sustrik yes
[14:23] sustrik no idea
[14:23] pieterh didn't you read the guide yet?
[14:24] reuben haha
[14:24] sustrik if you know about all the active peers
[14:24] pieterh no, srsly
[14:24] sustrik you can do load-balancing in the user space
[14:24] pieterh did you read the guide, where we actually build patterns in userspace, using router?
[14:24] sustrik some of it
[14:24] pieterh there are 10-20 clear use cases there
[14:24] pieterh including *exhaustive* explanation of why router-to-router doesn't make sense
[14:24] pieterh except in one single usecase
[14:25] pieterh we've discussed this multiple times wrt identities
[14:25] sustrik can you give me a pointer?
[14:25] pieterh ok,
[14:25] pieterh a router cannot route until it gets an incoming message
[14:25] pieterh it is a server semantic
[14:25] pieterh it needs a client to talk to it first and say, "I'm here"
[14:25] sustrik that's why i am saying the connection/disconnection notification would have to be added
[14:25] pieterh no difference
[14:26] pieterh a router doesn't know addresses of inexistent clients up front
[14:26] pieterh the only way
[14:26] pieterh is if you use explicit identities
[14:26] sustrik that can be added to the generalised router socket
[14:26] pieterh and the only sane way I found was to make them ipaddress:port
[14:26] sustrik connection happend => user gets notification message containing connection ID
[14:26] pieterh so you want to mix router/dealer semantics together
[14:26] sustrik happens*
[14:26] pieterh into a kind of "whatever" socket
[14:26] sustrik yes
[14:27] pieterh sure
[14:27] pieterh if you're going to do that
[14:27] sustrik i can
[14:27] sustrik if you find it useful
[14:27] pieterh then you can make fair queuing, etc. comfigurable
[14:27] sustrik whatever you want
[14:27] sustrik basically, it would allow 0mq to be used as a dumb networking framework
[14:27] sustrik such as ACE
[14:27] pieterh well, to be honest, the problem isn't lack of power
[14:27] pieterh the problem is complexity
[14:28] pieterh it is not worth just adding functionality if the semantics become too complex to use
[14:28] sustrik that's the question
[14:28] pieterh so, what I'd suggest is a SERVER / CLIENT semantic
[14:28] pieterh very explicit
[14:28] pieterh client can connect to exactly one server
[14:29] pieterh server is ROUTER plus connect/disconnect notifications
[14:29] sustrik right client/server pattern
[14:29] pieterh they do heartbeating and possibly some other stuff like credit-based flow control instead of HWMs
[14:29] pieterh I have a design for CBFC if you want it, but it's what TCP does, not very complex
[14:30] sustrik the problem with that is that that way i would have to implement every new pattern in 0mq
[14:30] sustrik no new patterns in user space
[14:30] pieterh what do you mean?
[14:30] pieterh all the high level patterns could be built on that
[14:30] sustrik if someone wanted some other pattern, it would still have to be implemetned insisde 0mq
[14:30] pieterh well, I don't see that
[14:31] pieterh you could implement all the existing patterns using server/client
[14:31] sustrik the one-connection-per-client disables whole class of patterns
[14:31] sustrik otherwise it's the same as i am proposing afaics
[14:32] pieterh the one connection per client could be configurable
[14:32] pieterh you need this in most simple cases to avoid utter confusion
[14:32] pieterh but in fact the dealer semantics are most accurate afaics
[14:33] sustrik ok, let me explain how i image generic router socket can work
[14:33] pieterh go for it
[14:33] sustrik you can do any number of binds or connects
[14:34] pieterh sure
[14:34] sustrik when connection is established, you get an notification with connection identity
[14:34] pieterh how do you get notifications? as messages?
[14:34] sustrik yes
[14:34] sustrik when disconnection happens, you get an notification
[14:34] sustrik when you send a message, you prepend it with identity
[14:34] pieterh sure
[14:35] sustrik when you receive a message it's prepended by identity
[14:35] sustrik that's it
[14:35] pieterh so just like router except with notifications
[14:35] sustrik completely generic
[14:35] sustrik yes, something like that
[14:35] pieterh now you still need a peer to talk to
[14:35] sustrik same socket type
[14:35] pieterh clearly you cannot connect two of these sockets together
[14:35] sustrik why not?
[14:35] sustrik it's just a glorified TCP
[14:36] pieterh could you make the identity into a schemed string?
[14:36] sustrik whatever
[14:37] pieterh it's a neat design, I like it
[14:37] sustrik ok, i can do that
[14:37] sustrik that way the people with non-stadard use cases still have an option
[14:38] pieterh well, apart from pub-sub, about 90% of use cases are non-standard afaics
[14:38] sustrik possibly
[14:38] pieterh there is pub-sub, pair for inproc pipes, and then req-rep for naive apps, and router-based patterns for everything else
[14:39] pieterh which works fine
[14:39] sustrik it would allow people to build alternative patterns on top of 0mq
[14:39] sustrik if some of them prove useful
[14:39] sustrik more effort can be spent in incorporating them directly into 0mq
[14:39] pieterh :-) we've been doing this (allowed or not) for about 12 months...
[14:39] pieterh since router got documented
[14:40] sustrik the problem at the time was that it was actually part of req/rep
[14:40] pieterh not really a problem, it worked fine
[14:40] sustrik it allowed applications to break req/rep semantics
[14:40] pieterh i didn't see that bug report
[14:40] pieterh is there a test case?
[14:42] sustrik easy, i can create some
[14:42] sustrik anyway, the problem is breaking the encapsulation
[14:43] sustrik back to generic routing: i can write the code
[14:43] sustrik the schemed identities are a bit of problem (testing on different platforms etc.)
[14:44] sustrik however, i can provide an extension point for transports
[14:44] sustrik to report the schemed identity
[14:44] sustrik so that it can be added easily
[14:45] pieterh what do you consider as a 'disconnection'?
[14:46] pieterh there is TCP disconnect, bound peer treating other as 'gone away', connected peer treating other as 'gone away'
[14:46] sustrik true
[14:47] pieterh the actual requirements are:
[14:47] sustrik explicit identities and thus the notion of session surviving disconnection would have to be removed before creating the new socket type
[14:47] pieterh - client mostly never wants a disconnect, but in some cases may want, via heartbeating
[14:47] pieterh - server wants to define a configurable timeout after which it can kick dead clients
[14:48] pieterh ok, so it's TCP connection (or whatever the transport layer is)
[14:48] sustrik ack, way for explicit disconnection would have to be added
[14:48] pieterh indeed
[14:48] pieterh forget heartbeats, that can be added on top
[14:49] sustrik yes
[14:49] pieterh so I assume when doing an outbound connect, client would get a 'connected' message when that succeeded?
[14:49] sustrik yes
[14:50] sustrik one more question: what should be the behaviour on HWM?
[14:50] sustrik block, i presume
[14:50] pieterh hang on, one other thought
[14:50] pieterh for notifications and commands
[14:51] pieterh we probably need a reserved identity that is actually the socket itself
[14:51] pieterh sys://socket or whatever
[14:51] sustrik good question
[14:51] pieterh working out-of-band would also be doable
[14:51] sustrik i would rather pass them directly via the socket
[14:52] sustrik but that requires some thinking about how to distinguish commands from data
[14:52] pieterh I mean, frame 0 = "sys://socket", frame 1 = "CLOSE", frame 2 = "identity of peer"
[14:52] pieterh that could go in both directions
[14:52] pieterh close command same as close notification
[14:52] pieterh open command? for consistency? bleh...
[14:53] sustrik it'll operate on hop-by-hop level anyway
[14:53] sustrik labels/commands are visible at that layer anyway
[14:54] sustrik e.g. "subscribe" command
[14:54] sustrik you can recv() it from XPUB socket
[14:54] sustrik and forward it to XSUB socket
[14:54] pieterh is this documented somewhere?
[14:55] sustrik zmq_socket(3)
[14:55] sustrik also in form of whitepaper: 250bpm.com/pubsub
[14:55] sustrik it should probably get into the guide later on
[14:56] pieterh the guide will always cover the stable version... so eventually it'll get there
[14:57] sustrik ack
[14:58] pieterh ok, so you use a simple binary framing for commands...
[14:58] sustrik possibly
[14:58] pieterh if router socket always sends/recvs address+message
[14:58] pieterh then you need an address that means "the socket itself not some other peer"
[14:58] sustrik or labal with no data = command ?
[14:58] sustrik label
[14:59] pieterh then you can use the same approach for commands in the message, e.g. a byte 1 for connect, byte 2 for disconnect
[14:59] pieterh labels aren't visible to the application, are they?
[14:59] sustrik for example
[14:59] sustrik they are visible on hop-by-hop layer
[14:59] sustrik invisible on end-to-end layer
[14:59] pieterh you said we'd use this socket at end-to-end as well as hop-by-hop
[15:00] pieterh you are mixing a lot of concepts here, it feels like mayonnaise
[15:00] pieterh with ketchup
[15:00] sustrik no, what i mean was that the "generic" socket can be a hop-by-hop concept
[15:00] pieterh that has no meaning to me
[15:00] sustrik you can build whatever end-to-end layer on top
[15:00] sustrik the goal is to allow custom patterns afterall
[15:01] sustrik on top = in user space
[15:01] pieterh there is a scaling disconnect here
[15:01] sustrik sorry?
[15:01] pieterh the only problems we're solving with router are where hop-by-hop == end-to-end
[15:02] sustrik ok
[15:02] pieterh so you are making a distinction that makes no sense to the problem
[15:02] pieterh I'd say we're several years away from understanding how e.g. to do multiple hops for the more complex patterns
[15:02] pieterh analogous to doing federation in AMQP, that took years to figure out
[15:03] sustrik at least the design it's future proof, you can use generic socket to create even complex scalable patterns on top of 0mq
[15:03] sustrik but you don't have to
[15:03] pieterh power isn't profitable
[15:03] pieterh simplicity is
[15:03] pieterh simplicity at scale is magic
[15:04] pieterh power at scale is just a mess
[15:04] sustrik if you have any idea how the end-to-end layer on top of genric socket should look like
[15:04] sustrik propose it
[15:04] pieterh I'd say you construct that on top
[15:04] pieterh out of multiple hops
[15:05] pieterh we do that already, e.g. where a client is actually a server for a set of local workers
[15:05] pieterh it's just not one pattern, it's different patterns, interconnected
[15:06] pieterh anyhow, this is mostly philosophy
[15:06] sustrik anyway, i'll implement the hop-by-hop layer first
[15:06] sustrik then we can discuss what should be on top of iot
[15:06] pieterh "label with no data part means a command" is just confusing several things IMO
[15:06] sustrik it
[15:06] sustrik possibly
[15:06] pieterh command = simplest possible way for application to pass connection meta data to/from socket
[15:07] sustrik i have no opinion on framing commands atm
[15:07] sustrik does "subscribe" command fit the definition?
[15:07] pieterh so for HWM
[15:07] pieterh it should block, yes
[15:07] sustrik ack
[15:08] pieterh i think the rest of ROUTER semantics are accurate, e.g. sending to non-connection means 'discard'
[15:08] sustrik as you wish
[15:08] sustrik it can report an error though
[15:08] sustrik no idea what's better
[15:08] pieterh that would be nicer, yes
[15:09] pieterh sending to a non-existent peer is bad enough to possibly warrant an assertion even
[15:09] sustrik ok
[15:09] pieterh an error would be good
[15:09] sustrik actually, error
[15:09] pieterh how about naming the socket type?
[15:10] sustrik dunno
[15:10] sustrik ZMQ_GENERIC
[15:10] sustrik whatever
[15:10] pieterh well, it could be an evolution of ROUTER
[15:10] pieterh which replaces ROUTER and DEALER
[15:10] pieterh a REALER
[15:10] pieterh or DOUTER
[15:10] sustrik the problem is backward compatibility
[15:11] pieterh zmq_setsockopt
[15:11] pieterh but ok
[15:11] sustrik ZMQ_RAW
[15:11] pieterh well, it's not raw
[15:11] pieterh actually, PEER would work here IMO
[15:12] sustrik maybe
[15:12] pieterh if this works, we can eventually kill router & dealer
[15:13] sustrik that would be nice, but would break backward comaptibility :|
[15:13] sustrik anyway, we can figure that out later on
[15:13] pieterh well, "kill" in the sense of sending to a far island and pretending we never heard of them...
[15:13] sustrik right
[15:14] sustrik last question: is usign generic socket any simpler that using raw TCP?
[15:14] sustrik using*
[15:14] pieterh so, suggestion... why not make this as an extension of router, enabled by setsockopt
[15:14] pieterh oh, far far far simpler
[15:14] sustrik hm
[15:14] sustrik message delimitation
[15:14] pieterh there's about 5,000 lines of code saved
[15:14] sustrik what else?
[15:15] pieterh you mean, apart from framing, multiparts, async i/o, error handling, portability, epoll/kpoll/select, ?
[15:16] sustrik framing, multipart and portability
[15:16] pieterh apart from working over multiple transports invisibly
[15:16] sustrik bingo
[15:16] pieterh performance
[15:16] sustrik muliple transports is a killer feature
[15:16] sustrik hm
[15:16] pieterh yes
[15:16] sustrik actually, it won't work over multicast
[15:17] pieterh but so it the rest, I know cause I'm reimplementing a lot of it in VTX
[15:17] pieterh *is the rest
[15:17] sustrik true
[15:17] pieterh asynchronous connects
[15:18] pieterh reconnects, from the client out to the server
[15:18] sustrik yes
[15:18] sustrik there's value in it
[15:18] sustrik you are right
[15:19] pieterh i can actually tell you how many days 0MQ saves in a realistic TCP client-to-server case
[15:19] pieterh without TCP, minimal cost to make a scalable client-server design is about 3 months
[15:19] pieterh without *0MQ, sorrt
[15:20] pieterh with 0MQ, about 4 hours
[15:20] sustrik in C, right?
[15:20] pieterh in C, on one single platform (Linux)
[15:20] pieterh if you were copying a lot of code, make that 2-3 weeks
[15:20] sustrik ok, makes sense
[15:21] pieterh 0MQ still has big holes but we assume these can be fixed without hurting the apps
[15:21] pieterh e.g. paranoia about ... oh, I have a great idea...
[15:21] pieterh we could make just this socket paranoid, to start with
[15:21] sustrik in what sense?
[15:21] pieterh max clients, max message size, max queue memory used, etc.
[15:21] pieterh I'm not sure if that's easier on a per-socket basis or globally
[15:22] sustrik globally is better
[15:22] pieterh ok, I didn't say anything
[15:22] sustrik the infrastructure for that is shared
[15:22] pieterh also the real magic with 0MQ sockets over TCP is that you use them both for talking to the outside world
[15:22] pieterh and for internal multithreading
[15:23] pieterh that's the real killer feature IMO
[15:23] sustrik have you seen that use case in the wild?
[15:24] pieterh every single advanced use case is multithreaded
[15:24] pieterh in the Guide too
[15:24] pieterh it's inevitable
[15:24] pieterh and it would be impossible if we had one loop for inter-thread events and one for inter-process events
[15:24] sustrik ah, that way
[15:24] pieterh s/impossible/horribly painful/
[15:24] sustrik right
[15:25] pieterh coffee, brb
[15:37] ssi I have a pretty good internal multithreading application with 0MQ
[15:37] ssi it's kind of like a pipeline pattern, but with LRU queue for flow control
[15:46] sustrik what i originally imagined people moving pieces of functionality from inside the process (threads) into separate applications
[15:46] sustrik what pieter had in mind was that in-process and remote logic can be monitored in a single event loop
[15:47] sustrik the in-process transport itself is definitely useful as such though
[16:10] cremes sustrik: i use inproc for inter-thread communications in my programs instead of mutexes
[16:10] cremes works great
[16:11] cremes and its conceptually much simpler for me to understand
[16:11] sustrik cremes: yes, understood
[16:18] pieterh sustrik: we have both use cases
[16:18] pieterh in CZMQ I call these 'detached' and 'attached' threads
[16:19] pieterh detached threads use their own context, and communicate over IPC or TCP or PGM
[16:19] pieterh attached threads share the parent context and talk to the parent thread over an inproc 'pipe' (a pair of PAIR/inproc sockets CZMQ creates automatically)
[16:19] pieterh this is cleaner than saying things like "create one context per process"
[16:34] sustrik pieterh: what we are going to do about the SIGPIPE thing
[16:34] sustrik i have nowhere to test it
[16:34] pieterh ah, right, got a little distracted
[16:34] sustrik do you have a mac?
[16:34] pieterh we have a user with a reproducible case
[16:35] pieterh Bill Hathaway
[16:35] pieterh I'll make a patch, send it to him, ask him to try it
[16:35] sustrik ok
[16:35] pieterh there's one place we set options on new sockets, right?
[16:36] sustrik you mean tcp sockets?
[16:36] sustrik yes
[16:36] sustrik tcp_socket.cpp
[16:36] sustrik :)
[16:36] pieterh goodo
[16:41] pieterh sustrik: ok, patched, we'll see what Bill says.
[18:24] wailupe2k So have a queue that pulls from a zmq and broadcasts to another, and periodically after a few million events it will display Assertion failed: nbytes == sizeof (command_t) (mailbox.cpp:194), The publisher is perl, and the subscriber is node.js.... Any ideas?
[18:36] ssi hrm this is annoying... trying to map my workers by socket identity, and the socket identity is coming back empty
[18:36] ssi it's after the socket's been created, connected, and a message sent
[18:40] ssi hrm, nm
[18:41] guido_g http://api.zeromq.org/2-1:zmq-getsockopt <- see ZMQ_IDENTITY
[18:41] ssi it was some kind of goofiness in constructing a string from the byte[] that identity returns
[18:41] ssi (jzmq)
[18:41] guido_g ahh ok
[18:42] guido_g oh twitter says pieterh is happy with amqp/1.0 ]:->
[18:43] pieterh guido_g: happy in the sense that watching other peoples' disasters makes us all feel more alive, yes
[18:43] ssi hahah
[18:43] guido_g 13 states
[18:44] guido_g sounds like a bad movie :)
[18:46] pieterh A bad Czech movie from 1993, "13 States of Connection"
[18:46] pieterh Starring Ivan N. Terpryze
[18:47] guido_g http://www.imdb.com/title/tt1473773/
[18:47] guido_g Drama 2010
[18:47] guido_g Not yet released
[18:47] guido_g so true :)
[18:47] pieterh sweet lord it actually exists?!
[18:47] pieterh what are the chances of that?
[18:48] guido_g Quotes
[18:48] guido_g Matthew O'Connor: We need someone who's professional and reliable - 'cos we're not.
[18:48] guido_g omg
[19:54] cremes wailupe2k: what version of libzmq and what OS?
[19:56] wailupe2k BSD, ZMQ 2.1.7
[19:57] sustrik that's the old problem with socketpair buffer limits
[19:57] sustrik is solved in 3.0
[19:57] wailupe2k hum
[19:57] wailupe2k do you know of a BSD port for 3?
[19:59] sustrik you can just build it
[19:59] sustrik the problem is rather with bindings
[19:59] sustrik 3.0 was released day before yesterday
[19:59] wailupe2k ah,
[19:59] sustrik and binding haven't caught up yet
[19:59] reuben 3.0 was released?
[19:59] wailupe2k this is on a prod sys :(
[20:00] sustrik well, i can point you to the patch
[20:00] sustrik you can try to backpoty it to 2.1.7
[20:00] sustrik backport*
[20:00] wailupe2k cool
[20:01] sustrik wait a sec
[20:02] sustrik this is the patch: https://github.com/zeromq/libzmq/commit/7c0c79812075459765440ca26bad56f4f7ddbe52
[20:03] wailupe2k hum,,, lots got cut out there.....
[20:03] wailupe2k Thanks though!
[20:03] wailupe2k sustrik++
[20:11] cremes wailupe2k: look up the faq and tuning guides on zeromq.org
[20:11] cremes wailupe2k: and adjust your system socket buffers
[20:12] wailupe2k cremes: oh, cool I'll check that out right now
[20:12] cremes 0mq 2.x uses a socketpair internally for signaling so the small bsd defaults cause problems like what you are seeing
[20:16] pieterh sustrik: do you have a Jira issue for that pipes problem?
[20:16] pieterh I'm trying a backport right now
[20:16] sustrik hm, let me see
[20:22] sustrik probably not
[20:22] sustrik i can't find it
[20:23] pieterh perhaps https://zeromq.jira.com/browse/LIBZMQ-166?
[20:23] mikko php extension should be up to date with 3.0 soonish
[20:23] sustrik it's a mix of several issues
[20:23] sustrik the third one seems to fit
[20:23] sustrik mikko: nice
[20:24] pieterh mikko: you mean with 3.0 support?
[20:24] mikko yes
[20:24] mikko should work with both
[20:24] pieterh mikko: very nice
[20:25] pieterh sustrik: it'll have to do then, I need an issue for any change in 2.1...
[20:25] sustrik sure, create one
[20:25] xristos can someone take a look at this: http://paste.lisp.org/display/123275
[20:25] xristos it is very short
[20:25] sustrik i guess it was never recorded in the bug tracker
[20:25] pieterh do we have a test case for this? does the shutdown stress test do it ?
[20:25] xristos i start a server that binds to a pub socket
[20:26] xristos then start 1000 clients that subscribe to it
[20:26] xristos most of the time i get # Assertion failed: new_sndbuf > old_sndbuf (mailbox.cpp:183)
[20:26] sustrik pieterh: try shutdown_stress + increase number of parallel threads (THREAD_COUNT)
[20:26] xristos sometimes it works
[20:26] pieterh sustrik: is xristos' example not the socketpair exhaustion?
[20:26] sustrik basically, to reproduce it you need an application thread that is evented but doesn't process the events
[20:27] sustrik so, for example you can create a socket, bind it and never touch it again
[20:27] sustrik then create a destroy peers quickly
[20:27] pieterh sustrik: is this is? # Assertion failed: new_sndbuf > old_sndbuf (mailbox.cpp:183)
[20:27] sustrik yes
[20:27] pieterh lol
[20:28] pieterh xristos: we're actually looking for you and your test case... welcome!
[20:28] xristos cool
[20:28] xristos i use 256 clients in production
[20:28] xristos and it works fine
[20:28] pieterh literally, the line before you posted was asking if we had such a test case
[20:28] xristos but if i could pump them up to 1000+
[20:28] pieterh what language is this?
[20:29] xristos python
[20:29] pieterh xristos, can I ask you to create an issue for this, with your test case, at https://zeromq.jira.com
[20:29] pieterh then when/if we have a fix I'll ask you to test it for us
[20:30] xristos http://paste.lisp.org/display/123275#2
[20:30] xristos this should be better
[20:30] xristos sure
[20:30] pieterh actually a plain C testcase would be best
[20:30] pieterh if I can I'll make one, based on your Python code
[20:31] pieterh thanks for coming to IRC, you can't know what perfect timing that was
[20:31] sustrik pieterh: should be easy to reproduce, you can use inproc for that
[20:31] sustrik s = socket (ZMQ_PUB);
[20:31] sustrik s.bind ("inproc://a")'
[20:32] sustrik while (true) {
[20:32] sustrik s2 = socket (ZMQ_SUB);
[20:32] sustrik s2.connect ("inproc://a");
[20:32] sustrik s2.close ();
[20:32] sustrik }
[20:33] pieterh I'll try it
[20:37] pieterh sustrik: ack, that kills it, after 430 sockets
[20:38] pieterh xristos: I've got a native C test case working
[20:38] xristos ok
[20:39] xristos it's only for PUB/SUB from what i see here
[20:39] xristos PULL/PUSH works fine up to 5k
[20:39] xristos that i've tried
[20:55] xristos pieterh: will you file the issue yourself?
[20:55] pieterh sustrik: well, I have a test case but it dies at 510 sockets with 'too many open files' (after the change)
[20:56] pieterh xristos: I'd prefer if you file the issue, and test the eventual change
[20:56] pieterh it is better that way
[21:02] xristos filed
[21:02] pieterh cool!
[21:02] pieterh well, I think I have a fix for it, seems to work in 2.1.x
[21:03] pieterh xristos: are you up to building 2.1.x from git master?
[21:07] xristos i will, tomorrow, i'm leaving work atm
[21:10] pieterh ok, I'll post some comments the jira issue
[21:23] benwaine hello
[21:24] pieterh benwaine: hi
[21:25] benwaine I've been reading through 'the guide' and have a question. Early on it mentions 'if you are opening lots of sockets all the time your probably doing it wrong' (paraphrase)
[21:26] benwaine I wanted to ask - i'm thinking of having a fixed queue which messages are pushed into.
[21:27] benwaine A user would enter a term into the ui and on the server side a socket would be connected to the queue and that once message would be sent.
[21:27] benwaine thats a lot of connects. Is this wrong?
[21:28] pieterh benwaine: you'
[21:28] pieterh you'd probably want to keep server processes running
[21:28] pieterh and only open / connect sockets the first time
[21:29] pieterh if only because opening / connecting a socket over TCP takes a small while and that'll add latency to your UI for nothing
[21:31] benwaine hmm ok. I'm writting this in PHP. Can you suggest how I should send the message to the queue without using a connect. Some kind of wrapper round the push socket?
[21:33] pieterh doesn't php let you keep server processes alive?
[21:33] pieterh this is like database handles, I guess
[21:34] benwaine I could run a script in a big loop (this is what i'm doing for the actual queue) but i think the connect would have to be on a per request basis. I'm still knew to this thought. Open to suggestions :)
[21:34] pieterh I don't know PHP in this regard
[21:35] pieterh my advice is to open sockets only when you need to, cache them, reuse them
[21:35] pieterh that'll give you best performance
[21:35] pieterh however if you need to get started simply, just open them on demand, close when done
[21:35] benwaine ok. I'll have a play round and keep reading the guide. Thanks for your time.
[21:35] pieterh np
[21:36] pieterh sustrik: ok, backported the socketpairs changes to 2.1...
[21:36] pieterh it was quite a lot of changes, more than that single commit
[21:36] dontfudge Hey all. I'm trying to get zmq/lua/zmsg going. I can't seem to find anywhere czmq for lua. Any clues?
[21:36] pieterh dontfudge: czmq is for C, not Lua
[21:37] dontfudge There's a example of Asynchronous client-server in Lua on zguide that requires zmsg, but I can't figure out where to get it.
[21:38] ianbarber benwaine: yeah, you'd likely create and drop a socket each time in that model with php. The overhead is pretty low.
[21:38] ianbarber dontfudge: it's usually in the same examples directory
[21:39] benwaine thanks Ian
[21:39] dontfudge It's not found in the git, but is here http://zguide.zeromq.org/lua:asyncsrv
[21:39] ianbarber https://github.com/imatix/zguide/blob/master/examples/Lua/zmsg.lua
[21:39] dontfudge yeah, saw that - but that's just an example on how to use zmsg in lua
[21:40] ianbarber no, that's it
[21:40] ianbarber zmsg started life as a helper for the guide examples in C, it got ported along with those examples
[21:40] dontfudge <sheepish grin> ok thanks
[21:40] ianbarber :)
[21:40] pieterh ianbarber: thanks
[21:41] ianbarber :)
[21:41] ianbarber what are you porting down?
[21:41] ianbarber the protocol bits for 2.2?
[21:41] pieterh ianbarber: removing the socketpair limitations
[21:41] ianbarber ah, cool
[21:42] pieterh surprisingly it all seems to work after I've been at it with a hammer
[21:43] ianbarber you'll be a converted C++er by the end of the year :)
[21:43] pieterh ianbarber: sigh, it's looking that way... :-(
[21:44] pieterh ianbarber: does the notion of using 2.2 as a stepping stone between 2.1 and 3.0 seem reasonable?
[21:45] ianbarber yeah, i think it's good, means 3.0 can stay clear of b/c code, and people can bridge their infrastructure
[21:45] pieterh good
[21:45] pieterh ok, I'm off for the weekend, be back Monday
[21:46] pieterh meetup in Geneva on Saturday, in case anyone's in that zone
[21:47] ianbarber cool, have a good one, say hi to alvaro
[21:48] pieterh for sure!
[22:32] gdan has anyone tested to see which is faster: inproc or ipc?
[22:50] cloudhead gdan: I'd assume inproc is faster, seeing as it's limited to a single process
[22:55] whack cloudhead_: and wouldn't copy from process -> kernel and back again, probably also less syscalls
[22:56] whack but at the point where your maxing out a tcp/inproc/ipc thing you're probably hitting limits on cpu cache misses, etc
[22:57] cloudhead yea ipc will have some overhead