I think there's a problem with boot features, and I think we don't see it with the old style assemblies because the bundles are in fact all listed in the startup.properties as well as as boot features.
There are 3 bundles in other parts of the startup.properties that depend on Apache Karaf :: Management (3.0.0.SNAPSHOT): [ 14] [Installed ] [ ] [ 30] Apache Karaf :: Diagnostic :: Management (3.0.0.SNAPSHOT) [ 16] [Installed ] [ ] [ 30] Apache Karaf :: Features :: Management (3.0.0.SNAPSHOT) [ 19] [Installed ] [ ] [ 30] Apache Karaf :: Admin :: Management (3.0.0.SNAPSHOT) Even if I put the feature core bundle at start level 25 it installs the boot feature asynchronously so the management bundle isn't installed by the time these 3 bundles need it. I'm not 100% sure but I think this problem still occurs if I make the feature service install the boot features synchronously. I'm surprised but can't debug through it tonight. One possible fix might be to move more of this into boot features so there are no dependencies from startup.properties-started-bundles on boot feature bundles. This is a pretty big change to the structure of the minimal server. Also the boot features don't appear to be started by default unless I put start='true' into each bundle. Any suggestions? thanks david jencks On Apr 17, 2011, at 12:23 PM, David Jencks wrote: > Hi Guillaume, > > I think you are suggesting: > > 1. combine local-repo and system under a suitable name such as system or > repository. I'll use system for now. > > 2. have as little as possible in startup.properties. For instance the > startup.properties for minimal and full servers should be the same > > 3. everything else in the basic minimal and full servers should be started > through bootFeatures. We want to ensure that all the required bundles are in > the (single) repo. > > 4. generally to create a custom server you'll add (or maybe also remove) more > bootFeatures and make sure the bundles are installed into the repo. > > This seems entirely reasonable to me. The only quibble I can think of at the > moment is that if everything is in startup.properties you can look at one > file and see everything that will be started in one place. > > I'm not sure how it works right now, but with this approach I think we'll > want to make sure that when the features service starts it installs all the > boot feature bundles immediately so the framework can use the start level > specified in the features and the order listed in bootFeatures won't matter. > > One situation this approach won't work for is if you need some bundles > started before the feature service itself. For instance if you want jaxb > 2.2 used everywhere you'd need to install the jaxb 2.2 system with a lower > start level than the feature service. So some custom servers may need to > alter the startup.properties compared to the minimal server. While in this > case I expect the jaxb 2.2 feature I'll need will be in its own feature repo > so it can be handled by just listing it as a compile scope dependency, so its > bundles will get added to startup.properties, this kind of situation > indicates to me that startupFeatures can still be useful. > > I'll see if I run into any problems switching the new-style assemblies to > this approach. > > thanks! > david jencks > > > On Apr 17, 2011, at 12:00 PM, Guillaume Nodet wrote: > >> Note that I've recently enhanced the startup mecahnism to not *always* >> try to install the bundles listed in the startup.properties. >> Previously, they were always reinstalled at startup time. >> >> First, I don't really see the need for local-repo and system. I >> think we should just merge them. >> >> There's really no need for bootFeatures + startupFeatures. I think we >> should only have one mechanism. The original idea was that the core >> bundles are started with the startup.properties and everthing else >> installed using features, mostly through bootFeatures. Now that we >> have a better maven plugin, the plugin is able to do both, but this is >> a bit confusing. Note that there is some differences though: >> * bundles listed startup.properties have to be in the system repo as >> the real mvn url handlers can't be used >> * manually editing the statup.properties is much more tedious than >> changing the bootFeatures imho >> * you can't put bundles using wrap url handler in the startup.properties >> The third point is specially problematic I think. One possible use >> case that isn't really addressed yet is the ability to create a >> configuration where some features depend on url handlers that are >> provisioned by other features. Or, said another way, a feature can >> only be *installed* if a dependent feature is *started*. I think >> moving things back in the startup.properties will make such >> dependencies even harder to manage. >> >> I think I would have rathe gone the opposite direction and remove as >> much as possible from the startup.properties to use the bootfeatures >> instead, but it may be just me. >> >> Just to understand the goal, what kind of benefit do you see by >> listing all bundles in the startup.properties ? I suppose we could >> get rid of the features service at runtime then, but apart from that, >> I'm not sure to see. >> >> On Sun, Apr 17, 2011 at 09:41, David Jencks <[email protected]> wrote: >>> Well, I think it may promote confusion, but there are now 3 options for >>> features in features repositories of maven runtime scope: >>> >>> startupFeatures. The bundles in these features will get installed into >>> system and listed in startup.properties >>> >>> bootFeatures: The bundles in these features will get installed into >>> local-repo and the feature names added to bootFeatures in the features cfg >>> file >>> >>> installedFeatures: The bundles in these features will get installed into >>> local-repo. >>> >>> For instance, >>> >>> <configuration> >>> <startupFeatures> >>> <feature>ssh</feature> >>> <feature>config</feature> >>> <feature>management</feature> >>> </startupFeatures> >>> </configuration> >>> >>> Note that except for startupFeatures the bundles are installed into >>> local-repo. I think system should only have bundles started from >>> startup.properties in it. >>> >>> thoughts? >>> >>> thanks >>> david jencks >>> >>> On Apr 16, 2011, at 12:50 PM, Johan Edstrom wrote: >>> >>>> For me it'd be the same logic as not "starting" everything from tomcats >>>> endorsed lib and >>>> just deploying WAR file, perhaps a little convoluted comparison, but the >>>> features additions, adding >>>> things and keeping "application" vs. features does make it simpler to >>>> manage as your features and deployments grow. >>>> >>>> >>>> On Apr 16, 2011, at 1:05 PM, David Jencks wrote: >>>> >>>>> If the same bundles all get started when you start the server, what is >>>>> the difference? I think you've "designed" the server you want by >>>>> including the bundles and indicating that you want them started at server >>>>> startup. How is it different whether the bundles are started from >>>>> startup.properties or the feature service looking at boot features? >>>>> >>>>> thanks >>>>> david jencks >>>>> >>>>> On Apr 16, 2011, at 11:48 AM, Johan Edstrom wrote: >>>>> >>>>>> I really think that the startup.properties is for infrastructure, >>>>>> bluperint etc. >>>>>> the startup features allows you do "design" your own distro. >>>>>> >>>>>> >>>>>> On Apr 16, 2011, at 12:42 PM, David Jencks wrote: >>>>>> >>>>>>> Hi JB, >>>>>>> >>>>>>> Thanks for the explanation, I think I've looked at things from only the >>>>>>> karaf-assembly point of view too long to see anything else :-) The way >>>>>>> karaf-assembly works now, all the bundles in system will be listed in >>>>>>> startup.properties and will start automatically. >>>>>>> >>>>>>> I think you are saying: >>>>>>> >>>>>>> -- if you include a feature in the boot feature list, you should make >>>>>>> sure all the bundles needed are in the server already. >>>>>>> >>>>>>> and, looking more closely at the add-features-to-repo mojo I see >>>>>>> there's a <features> configuration element that lets you specify >>>>>>> features to install the bundles for. >>>>>>> >>>>>>> I'm going to add that capability to the karaf-assembly packaging. >>>>>>> >>>>>>> There's another difference of style to resolve. I've set up the >>>>>>> karaf-assembly packaging so that whenever it installs a featue, it adds >>>>>>> the feature's bundles to startup.properties at the appropriate start >>>>>>> level. For "boot" features this means that you don't need to list them >>>>>>> in the features service configuration "boot features" since they will >>>>>>> be started via the startup.properties. On the other hand you can't >>>>>>> uninstall this feature so easily. Is this an important difference? Do >>>>>>> I need to make a way to install a feature's bundles but not add the >>>>>>> bundles to startup.properties? >>>>>>> >>>>>>> many thanks >>>>>>> david jencks >>>>>>> >>>>>>> >>>>>>> On Apr 16, 2011, at 12:12 AM, Jean-Baptiste Onofré wrote: >>>>>>> >>>>>>>> Hi David, >>>>>>>> >>>>>>>> I'm not sure to follow you. bootFeatures property defines which >>>>>>>> features are started at Karaf bootstrap time. It allows user to start >>>>>>>> a fresh Karaf instance with a set of Karaf features loaded and >>>>>>>> installed. >>>>>>>> >>>>>>>> For instance, in ServiceMix, we have: >>>>>>>> - nmr feature as boot feature and we ship all required jar (using >>>>>>>> features-add-to-repo goal) in the system OBR >>>>>>>> - camel-nmr, cxf-nmr, etc are part of the NMR features but not in the >>>>>>>> boot features set. As we consider this kind of features as optional, >>>>>>>> we don't ship the features jar in the system OBR >>>>>>>> >>>>>>>> Regards >>>>>>>> JB >>>>>>>> >>>>>>>> On 04/16/2011 08:32 AM, David Jencks wrote: >>>>>>>>> I am wondering why we have boot features in trunk. In particular I >>>>>>>>> think there is an inconsistency around the management feature which >>>>>>>>> is in the minimal boot feature list and whose jars are in the minimal >>>>>>>>> server assembly. On the other hand the ssh full boot feature doesn't >>>>>>>>> have all its jars in the full server. >>>>>>>>> >>>>>>>>> If a boot feature doesn't already have all its jars in the server, >>>>>>>>> won't this require internet access on initial startup? >>>>>>>>> >>>>>>>>> If all the jars for a boot feature are already installed in the >>>>>>>>> server, why call it a boot feature, why not just start it with the >>>>>>>>> rest of the jars? For this question I may be biased by looking at >>>>>>>>> the packaging based assemblies where the startup.properties is >>>>>>>>> constructed from the features that are installed into the server, so >>>>>>>>> it is at least as east to configure the assembly to just include the >>>>>>>>> jars as to configure a boot feature. >>>>>>>>> >>>>>>>>> Am I missing something? >>>>>>>>> >>>>>>>>> thanks >>>>>>>>> david jencks >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>> >>> >> >> >> >> -- >> Cheers, >> Guillaume Nodet >> ------------------------ >> Blog: http://gnodet.blogspot.com/ >> ------------------------ >> Open Source SOA >> http://fusesource.com >> >> Connect at CamelOne May 24-26 >> The Open Source Integration Conference >> http://camelone.com/ >
