|Subject:||Re: improving cross language maintainability|
|From:||Fraser Adams (fras...@blueyonder.co.uk)|
|Date:||Dec 20, 2013 8:43:00 am|
I've been following this thread with interest and I guess that Rob's comment
However I think that would require us to be organised differently with a recognition that the Engine and Messenger are conceptually separate (with Messenger depending on the Engine API, and having no more authority over defining/testing the engine API than any other client "
Struck some resonance.
I think perhaps the layering is OK if you're familiar with the code base and
perhaps it's more about "packaging" than layering, but FWIW coming into proton
quite "cold" it seems like messenger and engine are essentially "combined". Now
I know that's not the case now from the responses to my "how does it all hang
together" question of a few weeks ago, but in terms of packaging that's still
kind of the case.
So what I mean by that is that qpid::messaging and proton messenger are as far
as has been discussed with me "peer" APIs - both high level that can achieve
similar things albeit with different semantics, whereas engine is a lower level
Why is it in that case that the proton library is built that contains both
messenger and engine? OK, so it's convenient, but as far as I'm aware neither
qpid::messaging nor Ted's dispatch router actually use messenger at all, they
both use engine? That would suggest to me that engine and messenger probably
ought to be exposed as separate libraries? (and presumably there's a similar
position in Java where the JMS implementation is engine not messenger based -
though I've not looked at the jar packaging).
I guess (perhaps related) it just might be better if messenger was in a separate
source tree, which I'd agree might be a faff, but would clarify the hierarchical
rather than peer relationship between messenger and engine.
So I guess that my take is that Rafael is perfectly accurate when he says that
"I believe they are currently layered precisely the way you describe" that certainly seems to be the case at the logical level, but at the
"physical" level it's an awful lot less clear of the layering - certainly from
the perspective of a novice trying to navigate the code base.
I'd have to agree also with Rob about the horizontal scaling aspects of lack of
comments and documentation, particularly with respect to engine. I'm currently
given the examples and documentation around messenger (though as you know even
there I've had a fair few complications) my personal preference/familiarity is
however around the JMS/qpid::messaging API and I'd ultimately quite like to have
begin and I'd probably end up "reverse engineering" the work Gordon has done
with qpid::messaging and/or Rob and co. are doing with the new JMS
I'd be interested to hear from Gordon and Ted about how they went about building
capability on top of the engine API.
I hope I'm not coming across as over-critical - email is a bit rubbish at
conveying emotion :-) I guess that I just want to add my 2p from the perspective
of someone coming in new with no previous exposure to the code base, which is
hopefully a fairly objective position. Best regards, Frase
On 20/12/13 15:57, Rafael Schloming wrote:
On Fri, Dec 20, 2013 at 5:46 AM, Rob Godfrey <rob....@gmail.com>wrote:
Since I have pretty much given up on trying to work on proton, I'm not sure that my opinion counts for much anymore... however I suppose I'm a little curious how the JNI binding differs in "extra work" vs. PHP or Ruby or such. Similarly why the API distinction is "hasty" when again it should (modulo some language idiom changes) reflect the API that the C is presenting to Ruby/PHP/Python/etc.
That's a really good question. One obvious factor is that the JNI bindings cover the full surface of the engine API, whereas the PHP and Ruby bindings only cover the messenger API, the latter being much smaller. The other factor is that we have a different workflow with how we maintain the scripting language bindings. From a design perspective each of those bindings are more tightly coupled to the C code since that is the only implementation they use, but from a workflow perspective each one is actually more independent than the Java binding. When we add a feature to the C messenger implementation we don't try to add it simultaneously across all the bindings. It generally goes into python because that is how we test, and it goes into perl and ruby pretty quickly after that because Darryl is good about keeping those up to date, but, e.g., in the case of php, we haven't tried to keep the binding at full feature parity. Finally, the JNI binding is a bit of a different animal from the others in that it is not just performing a simple wrapping of the C API and directly exposing that to Java. It is adapting between the C API underneath it and the Java API above it.
As for the hastiness, perhaps that was a poor choice of words. The problem with the Java API/impl split is that the API was produced from an existing implementation after the fact in order to facilitate testing. As such it isn't well suited to be implemented by two independent implementations, particularly one that is JNI based.
To me the fact that APIs aren't defined but that instead the only way of knowing what a particular call should do seems to be 1) look at the code in C and, if it's not clear, then 2) ask Rafi; is why I no longer feel it productive to try to contribute. As such, it seems that this change is more about vertical scaling (making it easier for Rafi, and those sitting very near him, to work more productively), than horizontal scaling (allowing people who are not Rafi or with easy access to him to work on it).
Really there are two separate issues here. You're talking about conceptual barriers to contribution, and I'm talking about mechanical barriers. Both are barriers to contribution, and both prevent horizontal scaling. Fixing some of the mechanical barriers to contribution does happen to also improve vertical scaling, but as I said in my original post, I think the more important factor is making it easier/possible for people to make complete contributions.
FWIW, I agree we need lots of additional work around API definition and documentation. Having more time for that is also one of the reasons I would like to reduce some of the overhead involved in the current development process.
Overall I don't care very much about Messenger in Java (either pure Java or JNI), but I do care about the Engine. When the Proton Engine was first discussed it was always said that the C code should be able to be called from within Java as an alternative to a pure Java impl. - has this changed?
I'm fine with that as a goal, however I can't say it's a priority for me. More importantly, as I've said before, I don't believe the use of JNI in production is at all compatible with the use of JNI for testing purposes. The current JNI binding exposes each part of the engine API in very fine detail. This what you want for testing, so you can exercise every part of the C API from Java. This is probably exactly the opposite of what you'd want for production use though. I expect you'd want to minimize the number of times you cross the C/Java boundary and just tie into the engine API at key points that are shifting a lot of bytes around.
Going back to the horizontal vs. vertical scaling... If the project were structured differently I would be very happy to contribute to the engine side of things in Java. However I think that would require us to be organised differently with a recognition that the Engine and Messenger are conceptually separate (with Messenger depending on the Engine API, and having no more authority over defining/testing the engine API than any other client - such as the C++ broker, ActiveMQ, or the upcoming JMS clent, etc.). As such I would rather see us fix the definition of the API (if you believe it to be "hasty") rather than simply try to remove any notion of their being an API which is distinct from the implementation.
I'm not sure exactly what you're asking for here with respect to Messenger and Engine. I believe they are currently layered precisely the way you describe. There certainly aren't any code level dependencies from Engine to Messenger. Are there specific ways that you think Messenger has influenced the Engine that are problematic? Is there something concrete you would like to see happen here?
As for the Java API, I'm not suggesting eliminating the notion of there being an API. What I'm suggesting is that what sits in the proton-api module is not actually an API that is well suited for multiple implementations, but is really a single implementation API. As such I would like to put it back into the proton package and focus on making it a good single implementation API. I don't believe that this step would detract at all from it being a well defined API, in fact I think it would make it clearer and more transparent as there are currently plenty of cases where the split obscures the design. I believe a distinct question is whether we need a multi-implementation API, and if so for what purpose, e.g. production use of a JNI based engine vs testing of the C based engine.