Re: Poll - updating HTTP to newer jetty

[Daniel Fagerstrom <da...@nada.kth.se>]

Rob Walker skrev:
> 
>> Have you started to work on this? I'm interested in working on 
>> implementing a Jetty 6 HTTP service.
> Haven't started yet but am hoping to get a chance to look in next week 
> or two.

I might be able to look at it sooner, in that case I post my findings 
immediately so that we avoid any double work.

>> As some background we are working on making Cocoon OSGi based. This 
>> far we use the Equinox framework and bundles as we rely heavily on 
>> declarative services, which hasn't been part of Felix until recently. 
>> The HTTPservice in Equinox that we have used until now take a 
>> fundamentalistic view on spec compliance and use servlet 2.1, while 
>> this have worked well during prototyping it is more or less unusable 
>> in a modern web framework. I have tried to switch to the current Felix 
>> HTTP service to be able to use servlet 2.3 but have got all kinds of 
>> classloading problems (unfortunately I'm not able to give any concise 
>> bug report as I don't have any clue about what happens). I have also 
>> tried the Oscar and Knopflerfish services who both works in my 
>> setting, but doesn't implement e.g. contextpath in a meaningful way.
> I think the problem here is going to be that to be a properly OSGi  
> usable bundle, will probably mean several things:
> 
> 1 - the base HTTP service we will make stick as close to OSGi compliance 
> as possible, which as you've found is a fairly "minimalist" Http service 
> in terms of more general web application facilities

I agree with that it should stick as close to OSGi compliance as 
possible. The only difference that I'd like at the moment is using 
Servlet 2.5 instead of Servlet 2.1 (actually I would be completely happy 
already with 2.3).

> 2 - I've heard that extensions and enhancements are in the wings to 
> improve what a standard OSGi HttpService offers - we'd look to implement 
> these as soon as available, but until then any richer "web frameworks" 
> will need to be by custom services and APIs. I believe a few others have 
> looked at providing extensions to offer richer, more standard web-app 
> capabilities on top of an OSGi service and was planning to send out a 
> poll once the Jetty6 work is done for who had what, and what interest 
> there was in standardizing an extended set of services

There is some work within OSGi on some kind of webapp service(s), 
unfortunately they insist on continuing old fashioned closed 
specification development so I have no idea what they are up to.

> I need to look back over the servlet JAR level discussion thread - 
> suffice to say, we'll try and balance spec compliance with the interest 
> for support as recent a level of servlet spec as possible.
>> Jetty 6 seem to be designed for embedding and doesn't seem to require 
>> context classloader tricks.
>>
> True in terms of embedding, it does which I'm very glad to see and shall 
> be even happier if it turns out simpler than the current model we use.
> 
> Classloading wise though, any integration will have to fit the standard 
> OSGi classloading model, which hopefully won't require classloader 
> tricks per se but still presents the potential for somewhat tricky 
> classloading errors at least during initial application implementation. 
> More recent versions of Felix have got a whole lot easier in terms of 
> meaningful messages to help trace these though.

I have no problem with the OSGi classloading model, quite the opposite. 
What I refered to is that it seem like at least the Oscar http service 
require that one resets the context classloader in some use cases where 
this isn't necessary for the Equinox http service. This complicates the 
use of it.

/Daniel

Re: Poll - updating HTTP to newer jetty

[Rob Walker <ro...@ascert.com>]

Daniel Fagerstrom wrote:
> I'm working on the implementation now. The architecture and the APIs 
> of Jetty6 has changed considerably since Jetty4 so it took some time 
> to figure out how to modify the current http.jetty implementation so 
> that it works with Jetty6. I'll put the code in  Jira as soon as I 
> have something working.
Good news and thanks! You've taken one off my list, which just leaves me 
mangen to worry about. More than happy to assist and or review code or 
design ideas/approaches if you want help or input on the approach.

There's a bugfix workaround submitted in JIRA to do with \* aliases for 
Http that I was going to apply to current Jetty4 version and see if was 
relevant for Jetty6. Don't worry if you don't have time for that as well 
- I can come back and look at it when I look at the new Jetty6 based 
Http code.

> Might be, I'm not certain. I needed to change the context class loader 
> while using the Oscar http service but not while using the Equinox one.
Context classloader issues are a bit of a pain in OSGi - in most cases 
it's a 3rd party JAR that needs it setting to do dynamic classloading 
(usually SAX parsers are the most prevalent case).

-- Rob

Re: Poll - updating HTTP to newer jetty

[Daniel Fagerstrom <da...@nada.kth.se>]

Rob Walker skrev:
> 
>> I might be able to look at it sooner, in that case I post my findings 
>> immediately so that we avoid any double work.
>>
> Cool - I was expecting to look at it in next few days, but have no 
> problem if you want to pitch in first.

I'm working on the implementation now. The architecture and the APIs of 
Jetty6 has changed considerably since Jetty4 so it took some time to 
figure out how to modify the current http.jetty implementation so that 
it works with Jetty6. I'll put the code in  Jira as soon as I have 
something working.

> My plan was to SVN copy the current http.jetty to a new http.jetty6 
> bundle initially, so that others use it and give feedback whilst keeping 
> the previous version intact for those wishing to stay on the stable 
> Jetty4 version. Also, this will track the mod deltas back to the 
> original too - so we won't lose any change information.

Have followed that.

> I'm also hoping to significantly clean-up the rather tricky way that we 
> need to wire into the standard Jetty stack,

That seem to be possible, Jetty6 seem to be a little bit more open for 
extension than Jetty4.

> and improve the way Jetty 
> logging is integrated with Felix logging.

I will probably wait with that.

>> I have no problem with the OSGi classloading model, quite the 
>> opposite. What I refered to is that it seem like at least the Oscar 
>> http service require that one resets the context classloader in some 
>> use cases where this isn't necessary for the Equinox http service. 
>> This complicates the use of it.
> 
>  From what I recall, the Oscar http service doesn't require use of 
> context classloader itself - there are numerous cases of third party 
> JARs which do e.g. various XML / SAX parsers etc. I'm guessing if you 
> found a need for your application to set a context classloader it was as 
> a result of a 3rd party JAR which was required by one your services 
> rather than the HTTP service itself.

Might be, I'm not certain. I needed to change the context class loader 
while using the Oscar http service but not while using the Equinox one.

/Daniel

Re: Poll - updating HTTP to newer jetty

[Rob Walker <ro...@ascert.com>]

> I might be able to look at it sooner, in that case I post my findings 
> immediately so that we avoid any double work.
>
Cool - I was expecting to look at it in next few days, but have no 
problem if you want to pitch in first.

My plan was to SVN copy the current http.jetty to a new http.jetty6 
bundle initially, so that others use it and give feedback whilst keeping 
the previous version intact for those wishing to stay on the stable 
Jetty4 version. Also, this will track the mod deltas back to the 
original too - so we won't lose any change information.

I'm also hoping to significantly clean-up the rather tricky way that we 
need to wire into the standard Jetty stack, and improve the way Jetty 
logging is integrated with Felix logging.

> I have no problem with the OSGi classloading model, quite the 
> opposite. What I refered to is that it seem like at least the Oscar 
> http service require that one resets the context classloader in some 
> use cases where this isn't necessary for the Equinox http service. 
> This complicates the use of it.

 From what I recall, the Oscar http service doesn't require use of 
context classloader itself - there are numerous cases of third party 
JARs which do e.g. various XML / SAX parsers etc. I'm guessing if you 
found a need for your application to set a context classloader it was as 
a result of a 3rd party JAR which was required by one your services 
rather than the HTTP service itself.

Regards

-- Rob


Ascert - Taking systems to the Edge
robw@ascert.com
+44 (0)20 7488 3470
www.ascert.com