Re: [INFO] Apache Karaf 5 will be internally based on OSGi R8

[Grzegorz Grzybek]

Quoting one statement of yours:

So for long running apps graal cost is wayy more than the runtime gain
> and I guess it is where Karaf 5 will sit, long running aggregated and
> unified apps (by providing a single admin interface for all kind of apps
> and not a different one for spring/spring-boot, microprofile, ee, osgi
> etc).
>

I agree. I still believe in "Java applications servers" (whatever the set
of standards - real or de-facto - is used). And I still believe in
long-running apps.

regards
Grzegorz Grzybek

czw., 25 mar 2021 o 10:33 Romain Manni-Bucau 
napisał(a):

> Le jeu. 25 mars 2021 à 09:35, Grzegorz Grzybek  a
> écrit :
>
> > Thanks Romain for the details! (see inline)
> >
> > czw., 25 mar 2021 o 08:31 Romain Manni-Bucau 
> > napisał(a):
> >
> > > Le jeu. 25 mars 2021 à 07:13, Grzegorz Grzybek 
> a
> > > écrit :
> > >
> > > > Good morning!
> > > >
> > > > śr., 24 mar 2021 o 19:57 Romain Manni-Bucau 
> > > > napisał(a):
> > > >
> > > > > in terms of arch yes, the key feature is to have a tree classloader
> > and
> > > > not
> > > > > a graph (drops all the build complexity of OSGi and enable scanning
> > > > > pluggability, yeah).
> > > > >
> > > >
> > > > Graph → Tree sounds like OSGi → JavaEE...
> > > > This can prevent user feature to install a bundle that overrides
> system
> > > > services I know that (without "134 Subsystem Service
> Specification"
> > > and
> > > > without hooks) effectively OSGi runtime is "flat" - every bundle wire
> > is
> > > > equal and resolution rules apply. Also every OSGi service is equal
> and
> > > > service rank is taken into account.
> > > >
> > >
> > > Yes and no, a service registration can still use @Priority or a SPI
> > method
> > > to be sorted, only thing it can prevent is to put conflicting deps in
> the
> > > same bootstrap classloader (that said these days OSGi is rarely used
> for
> > > that and since by design the bootstrap loader will be a single app - ie
> > > without any conflict at build time - it is actually sane).
> > >
> >
> > In JavaEE, a WAR can (mostly) configure some providers, so e.g.,
> > DocumentBuilderFactory may return WAR-specific instance. But it's not
> > possible to affect this service loading in other WARs.
> > In OSGi, a bundle can register some service that'll become the valid
> > service for remaining bundles.
> > So I understand that Karaf 5 keeps the OSGi philosophy here, right?
> >
>
> Yes and not, the small language trick is do you speak of bootstrap services
> or profile or app in Karaf 5.
> Bootstrap services can do whatever they want (ie same as OSGi in terms of
> impact even if technicaly it is not linked) but all other layers
> (profile+app) must stay static and almost immutable.
>
>
> >
> >
> > >
> > >
> > > >
> > > > I'm trying to imagine how "it’s powered by OSGi R8 (but you will see
> > that
> > > > it’s more an internal point)" works - is the "Level1: Karaf itself" a
> > > > graph-based layer of bundle classloaders, while applications are
> given
> > > > their own single classloader (kind of like WebSphere is (was?) based
> on
> > > > OSGi and WARs/EARs hand single classloader or like Wildfly/EAP that's
> > > > internally a graph of JBoss Modules, while WARs/EARs have single
> > > > classloader)?
> > > >
> > > > java.util.ServiceLoader is dynamic in nature and is a final (IMO) and
> > > quite
> > > > elegant discovery solution in tree-(ClassLoader)-based monoliths
> where
> > > you
> > > > "deploy" applications. And it's reflection based.
> > > >
> > >
> > > Last point does not have to be true, see some graalvm integrations for
> > > example, it is reflection less depends how you handle the build phase
> but
> > > being reflection "full" by default enables to keep the tooling
> (testing)
> > > working without breaking your IDE.
> > >
> >
> > Mind that I'm not very experienced with Graal/Quarkus, so my questions
> may
> > be invalid ;)
> >
>
> I was expecting it to come at some point - and btw we can note the fun
> thing that the big change is GraalVM but everybody speaks of Quarkus which
> is just a rebranding of already existing things, no technology jump by
> itself ;).
> My vision is that karaf 5 fulfills the microservices pitfalls and drawback
> by bringing back a well know and secure deployment alternative to all that.
> Indeed graal-ifying your app will make it save some memory, maybe some CPU
> cycle in some cases but if you optimize your java code you can get the same
> in terms of CPU cycles (and even faster in some cases).
> In terms of bootstrap you can same a few ms due to the classloading but not
> much more and CDS already solves part of it (at the cost of memory).
> So for long running apps graal cost is wayy more than the runtime gain
> and I guess it is where Karaf 5 will sit, long running aggregated and
> unified apps (by providing a single admin interface for all kind of apps
> and not a different one for spring/spring-boot, microprofile, ee, osgi
> etc).
>
> Hope it makes sens

Re: [INFO] Apache Karaf 5 will be internally based on OSGi R8

[Jean-Baptiste Onofre]

Just to be concrete about Karaf wording:

- Library: class loader at Karaf system level
- Service: loaded at Karaf system level
- Profile: class loader (not attended at karaf system level)
- Module: one class loader eventually with profile parent class loader and 
karaf system classloader

I hope it helps ;)

Regards
JB

> Le 25 mars 2021 à 10:33, Romain Manni-Bucau  a écrit :
> 
> Le jeu. 25 mars 2021 à 09:35, Grzegorz Grzybek  > a
> écrit :
> 
>> Thanks Romain for the details! (see inline)
>> 
>> czw., 25 mar 2021 o 08:31 Romain Manni-Bucau 
>> napisał(a):
>> 
>>> Le jeu. 25 mars 2021 à 07:13, Grzegorz Grzybek  a
>>> écrit :
>>> 
 Good morning!
 
 śr., 24 mar 2021 o 19:57 Romain Manni-Bucau 
 napisał(a):
 
> in terms of arch yes, the key feature is to have a tree classloader
>> and
 not
> a graph (drops all the build complexity of OSGi and enable scanning
> pluggability, yeah).
> 
 
 Graph → Tree sounds like OSGi → JavaEE...
 This can prevent user feature to install a bundle that overrides system
 services I know that (without "134 Subsystem Service Specification"
>>> and
 without hooks) effectively OSGi runtime is "flat" - every bundle wire
>> is
 equal and resolution rules apply. Also every OSGi service is equal and
 service rank is taken into account.
 
>>> 
>>> Yes and no, a service registration can still use @Priority or a SPI
>> method
>>> to be sorted, only thing it can prevent is to put conflicting deps in the
>>> same bootstrap classloader (that said these days OSGi is rarely used for
>>> that and since by design the bootstrap loader will be a single app - ie
>>> without any conflict at build time - it is actually sane).
>>> 
>> 
>> In JavaEE, a WAR can (mostly) configure some providers, so e.g.,
>> DocumentBuilderFactory may return WAR-specific instance. But it's not
>> possible to affect this service loading in other WARs.
>> In OSGi, a bundle can register some service that'll become the valid
>> service for remaining bundles.
>> So I understand that Karaf 5 keeps the OSGi philosophy here, right?
>> 
> 
> Yes and not, the small language trick is do you speak of bootstrap services
> or profile or app in Karaf 5.
> Bootstrap services can do whatever they want (ie same as OSGi in terms of
> impact even if technicaly it is not linked) but all other layers
> (profile+app) must stay static and almost immutable.
> 
> 
>> 
>> 
>>> 
>>> 
 
 I'm trying to imagine how "it’s powered by OSGi R8 (but you will see
>> that
 it’s more an internal point)" works - is the "Level1: Karaf itself" a
 graph-based layer of bundle classloaders, while applications are given
 their own single classloader (kind of like WebSphere is (was?) based on
 OSGi and WARs/EARs hand single classloader or like Wildfly/EAP that's
 internally a graph of JBoss Modules, while WARs/EARs have single
 classloader)?
 
 java.util.ServiceLoader is dynamic in nature and is a final (IMO) and
>>> quite
 elegant discovery solution in tree-(ClassLoader)-based monoliths where
>>> you
 "deploy" applications. And it's reflection based.
 
>>> 
>>> Last point does not have to be true, see some graalvm integrations for
>>> example, it is reflection less depends how you handle the build phase but
>>> being reflection "full" by default enables to keep the tooling (testing)
>>> working without breaking your IDE.
>>> 
>> 
>> Mind that I'm not very experienced with Graal/Quarkus, so my questions may
>> be invalid ;)
>> 
> 
> I was expecting it to come at some point - and btw we can note the fun
> thing that the big change is GraalVM but everybody speaks of Quarkus which
> is just a rebranding of already existing things, no technology jump by
> itself ;).
> My vision is that karaf 5 fulfills the microservices pitfalls and drawback
> by bringing back a well know and secure deployment alternative to all that.
> Indeed graal-ifying your app will make it save some memory, maybe some CPU
> cycle in some cases but if you optimize your java code you can get the same
> in terms of CPU cycles (and even faster in some cases).
> In terms of bootstrap you can same a few ms due to the classloading but not
> much more and CDS already solves part of it (at the cost of memory).
> So for long running apps graal cost is wayy more than the runtime gain
> and I guess it is where Karaf 5 will sit, long running aggregated and
> unified apps (by providing a single admin interface for all kind of apps
> and not a different one for spring/spring-boot, microprofile, ee, osgi etc).
> 
> Hope it makes sense and I'm not too far from what JB had in mind but this
> is where I see a looot of value for such a design.
> 
> 
>> 
>> 
>>> 
>>> 
 How about graal/quarkus?
 Let me be clear here - quarkus/graal/native approach is cool and makes
>>> Java
 great again™, but I know that "enteprise" still likes the idea of
 "appli

Re: [INFO] Apache Karaf 5 will be internally based on OSGi R8

[Romain Manni-Bucau]

Le jeu. 25 mars 2021 à 09:35, Grzegorz Grzybek  a
écrit :

> Thanks Romain for the details! (see inline)
>
> czw., 25 mar 2021 o 08:31 Romain Manni-Bucau 
> napisał(a):
>
> > Le jeu. 25 mars 2021 à 07:13, Grzegorz Grzybek  a
> > écrit :
> >
> > > Good morning!
> > >
> > > śr., 24 mar 2021 o 19:57 Romain Manni-Bucau 
> > > napisał(a):
> > >
> > > > in terms of arch yes, the key feature is to have a tree classloader
> and
> > > not
> > > > a graph (drops all the build complexity of OSGi and enable scanning
> > > > pluggability, yeah).
> > > >
> > >
> > > Graph → Tree sounds like OSGi → JavaEE...
> > > This can prevent user feature to install a bundle that overrides system
> > > services I know that (without "134 Subsystem Service Specification"
> > and
> > > without hooks) effectively OSGi runtime is "flat" - every bundle wire
> is
> > > equal and resolution rules apply. Also every OSGi service is equal and
> > > service rank is taken into account.
> > >
> >
> > Yes and no, a service registration can still use @Priority or a SPI
> method
> > to be sorted, only thing it can prevent is to put conflicting deps in the
> > same bootstrap classloader (that said these days OSGi is rarely used for
> > that and since by design the bootstrap loader will be a single app - ie
> > without any conflict at build time - it is actually sane).
> >
>
> In JavaEE, a WAR can (mostly) configure some providers, so e.g.,
> DocumentBuilderFactory may return WAR-specific instance. But it's not
> possible to affect this service loading in other WARs.
> In OSGi, a bundle can register some service that'll become the valid
> service for remaining bundles.
> So I understand that Karaf 5 keeps the OSGi philosophy here, right?
>

Yes and not, the small language trick is do you speak of bootstrap services
or profile or app in Karaf 5.
Bootstrap services can do whatever they want (ie same as OSGi in terms of
impact even if technicaly it is not linked) but all other layers
(profile+app) must stay static and almost immutable.


>
>
> >
> >
> > >
> > > I'm trying to imagine how "it’s powered by OSGi R8 (but you will see
> that
> > > it’s more an internal point)" works - is the "Level1: Karaf itself" a
> > > graph-based layer of bundle classloaders, while applications are given
> > > their own single classloader (kind of like WebSphere is (was?) based on
> > > OSGi and WARs/EARs hand single classloader or like Wildfly/EAP that's
> > > internally a graph of JBoss Modules, while WARs/EARs have single
> > > classloader)?
> > >
> > > java.util.ServiceLoader is dynamic in nature and is a final (IMO) and
> > quite
> > > elegant discovery solution in tree-(ClassLoader)-based monoliths where
> > you
> > > "deploy" applications. And it's reflection based.
> > >
> >
> > Last point does not have to be true, see some graalvm integrations for
> > example, it is reflection less depends how you handle the build phase but
> > being reflection "full" by default enables to keep the tooling (testing)
> > working without breaking your IDE.
> >
>
> Mind that I'm not very experienced with Graal/Quarkus, so my questions may
> be invalid ;)
>

I was expecting it to come at some point - and btw we can note the fun
thing that the big change is GraalVM but everybody speaks of Quarkus which
is just a rebranding of already existing things, no technology jump by
itself ;).
My vision is that karaf 5 fulfills the microservices pitfalls and drawback
by bringing back a well know and secure deployment alternative to all that.
Indeed graal-ifying your app will make it save some memory, maybe some CPU
cycle in some cases but if you optimize your java code you can get the same
in terms of CPU cycles (and even faster in some cases).
In terms of bootstrap you can same a few ms due to the classloading but not
much more and CDS already solves part of it (at the cost of memory).
So for long running apps graal cost is wayy more than the runtime gain
and I guess it is where Karaf 5 will sit, long running aggregated and
unified apps (by providing a single admin interface for all kind of apps
and not a different one for spring/spring-boot, microprofile, ee, osgi etc).

Hope it makes sense and I'm not too far from what JB had in mind but this
is where I see a looot of value for such a design.


>
>
> >
> >
> > > How about graal/quarkus?
> > > Let me be clear here - quarkus/graal/native approach is cool and makes
> > Java
> > > great again™, but I know that "enteprise" still likes the idea of
> > > "application servers", so I hope Karaf5 is NOT going to be
> > > "Kubernetes/OpenShift first" - long running processes with reflection
> and
> > > dynamic classloading are still relevant.
> > >
> >
> > Can you precise it there? Quarkus has two modes: JVM (where it is
> > equivalent to most microprofile servers without the standard/spec
> support)
> > and native mode (where arthur does the same closer to graalvm).
> > First mode does not need much but last one does not concern karaf 5 AFAIK
> 

Re: [INFO] Apache Karaf 5 will be internally based on OSGi R8

[Grzegorz Grzybek]

Thanks Romain for the details! (see inline)

czw., 25 mar 2021 o 08:31 Romain Manni-Bucau 
napisał(a):

> Le jeu. 25 mars 2021 à 07:13, Grzegorz Grzybek  a
> écrit :
>
> > Good morning!
> >
> > śr., 24 mar 2021 o 19:57 Romain Manni-Bucau 
> > napisał(a):
> >
> > > in terms of arch yes, the key feature is to have a tree classloader and
> > not
> > > a graph (drops all the build complexity of OSGi and enable scanning
> > > pluggability, yeah).
> > >
> >
> > Graph → Tree sounds like OSGi → JavaEE...
> > This can prevent user feature to install a bundle that overrides system
> > services I know that (without "134 Subsystem Service Specification"
> and
> > without hooks) effectively OSGi runtime is "flat" - every bundle wire is
> > equal and resolution rules apply. Also every OSGi service is equal and
> > service rank is taken into account.
> >
>
> Yes and no, a service registration can still use @Priority or a SPI method
> to be sorted, only thing it can prevent is to put conflicting deps in the
> same bootstrap classloader (that said these days OSGi is rarely used for
> that and since by design the bootstrap loader will be a single app - ie
> without any conflict at build time - it is actually sane).
>

In JavaEE, a WAR can (mostly) configure some providers, so e.g.,
DocumentBuilderFactory may return WAR-specific instance. But it's not
possible to affect this service loading in other WARs.
In OSGi, a bundle can register some service that'll become the valid
service for remaining bundles.
So I understand that Karaf 5 keeps the OSGi philosophy here, right?


>
>
> >
> > I'm trying to imagine how "it’s powered by OSGi R8 (but you will see that
> > it’s more an internal point)" works - is the "Level1: Karaf itself" a
> > graph-based layer of bundle classloaders, while applications are given
> > their own single classloader (kind of like WebSphere is (was?) based on
> > OSGi and WARs/EARs hand single classloader or like Wildfly/EAP that's
> > internally a graph of JBoss Modules, while WARs/EARs have single
> > classloader)?
> >
> > java.util.ServiceLoader is dynamic in nature and is a final (IMO) and
> quite
> > elegant discovery solution in tree-(ClassLoader)-based monoliths where
> you
> > "deploy" applications. And it's reflection based.
> >
>
> Last point does not have to be true, see some graalvm integrations for
> example, it is reflection less depends how you handle the build phase but
> being reflection "full" by default enables to keep the tooling (testing)
> working without breaking your IDE.
>

Mind that I'm not very experienced with Graal/Quarkus, so my questions may
be invalid ;)


>
>
> > How about graal/quarkus?
> > Let me be clear here - quarkus/graal/native approach is cool and makes
> Java
> > great again™, but I know that "enteprise" still likes the idea of
> > "application servers", so I hope Karaf5 is NOT going to be
> > "Kubernetes/OpenShift first" - long running processes with reflection and
> > dynamic classloading are still relevant.
> >
>
> Can you precise it there? Quarkus has two modes: JVM (where it is
> equivalent to most microprofile servers without the standard/spec support)
> and native mode (where arthur does the same closer to graalvm).
> First mode does not need much but last one does not concern karaf 5 AFAIK
> since spring-boot has its own graal integration, microprofile servers too
> (potentially EE ones too even if I didnt see one yet) and OSGi has its own
> through winegroewer so overall Karaf 5 sounds like the aggregator platform
> which would fallback on dropping it to be graal compliant (since you'll
> drop classloaders which makes all the power of the solution.
>

I imagine that Quarkus/Graal is designed mostly to develop apps that can
quickly start/stop and "application servers" is not the most desired goal
here.
And I was thinking about the native mode, where everything is mostly set up
at build time.



>
>
> >
> >
> > > In terms of service since the launcher is a monolith it has the key
> > > advantage to be able to scan all then dispatch so I guess we can just
> > have
> > > a ServiceLoader kind of SPI for "module service" impls and order them
> as
> > > needed. a ModuleService { setModuleServiceRegistry(Registry); } would
> > then
> > > do the trick probably, no need of fancy IoC for such low level
> framework
> > > IMHO.
> > >
> >
> > So clear distinguishing between "applications" and "server plugins" (with
> > e.g., replaceable Jackson as JSON provider) - am I interpreting your
> > statements correctly Romain?
> >
>
> For example yes even if I suspect the services should stick to very
> technical layers and isolated from the profile+app loaders so means jackson
> from the bootstrap loader shouldnt be usable in an app but you could
> configure it to leak (in the profile - ie the parent classes to use).
> Very generally services shouldnt leak but profiles will so a provider would
> sit in a profile loader IMHO.
> Services would be more about logging integratio

Re: [INFO] Apache Karaf 5 will be internally based on OSGi R8

[Romain Manni-Bucau]

Le jeu. 25 mars 2021 à 07:13, Grzegorz Grzybek  a
écrit :

> Good morning!
>
> śr., 24 mar 2021 o 19:57 Romain Manni-Bucau 
> napisał(a):
>
> > in terms of arch yes, the key feature is to have a tree classloader and
> not
> > a graph (drops all the build complexity of OSGi and enable scanning
> > pluggability, yeah).
> >
>
> Graph → Tree sounds like OSGi → JavaEE...
> This can prevent user feature to install a bundle that overrides system
> services I know that (without "134 Subsystem Service Specification" and
> without hooks) effectively OSGi runtime is "flat" - every bundle wire is
> equal and resolution rules apply. Also every OSGi service is equal and
> service rank is taken into account.
>

Yes and no, a service registration can still use @Priority or a SPI method
to be sorted, only thing it can prevent is to put conflicting deps in the
same bootstrap classloader (that said these days OSGi is rarely used for
that and since by design the bootstrap loader will be a single app - ie
without any conflict at build time - it is actually sane).


>
> I'm trying to imagine how "it’s powered by OSGi R8 (but you will see that
> it’s more an internal point)" works - is the "Level1: Karaf itself" a
> graph-based layer of bundle classloaders, while applications are given
> their own single classloader (kind of like WebSphere is (was?) based on
> OSGi and WARs/EARs hand single classloader or like Wildfly/EAP that's
> internally a graph of JBoss Modules, while WARs/EARs have single
> classloader)?
>
> java.util.ServiceLoader is dynamic in nature and is a final (IMO) and quite
> elegant discovery solution in tree-(ClassLoader)-based monoliths where you
> "deploy" applications. And it's reflection based.
>

Last point does not have to be true, see some graalvm integrations for
example, it is reflection less depends how you handle the build phase but
being reflection "full" by default enables to keep the tooling (testing)
working without breaking your IDE.


> How about graal/quarkus?
> Let me be clear here - quarkus/graal/native approach is cool and makes Java
> great again™, but I know that "enteprise" still likes the idea of
> "application servers", so I hope Karaf5 is NOT going to be
> "Kubernetes/OpenShift first" - long running processes with reflection and
> dynamic classloading are still relevant.
>

Can you precise it there? Quarkus has two modes: JVM (where it is
equivalent to most microprofile servers without the standard/spec support)
and native mode (where arthur does the same closer to graalvm).
First mode does not need much but last one does not concern karaf 5 AFAIK
since spring-boot has its own graal integration, microprofile servers too
(potentially EE ones too even if I didnt see one yet) and OSGi has its own
through winegroewer so overall Karaf 5 sounds like the aggregator platform
which would fallback on dropping it to be graal compliant (since you'll
drop classloaders which makes all the power of the solution.


>
>
> > In terms of service since the launcher is a monolith it has the key
> > advantage to be able to scan all then dispatch so I guess we can just
> have
> > a ServiceLoader kind of SPI for "module service" impls and order them as
> > needed. a ModuleService { setModuleServiceRegistry(Registry); } would
> then
> > do the trick probably, no need of fancy IoC for such low level framework
> > IMHO.
> >
>
> So clear distinguishing between "applications" and "server plugins" (with
> e.g., replaceable Jackson as JSON provider) - am I interpreting your
> statements correctly Romain?
>

For example yes even if I suspect the services should stick to very
technical layers and isolated from the profile+app loaders so means jackson
from the bootstrap loader shouldnt be usable in an app but you could
configure it to leak (in the profile - ie the parent classes to use).
Very generally services shouldnt leak but profiles will so a provider would
sit in a profile loader IMHO.
Services would be more about logging integration, http integration etc but
wouldnt leak as such but as a karaf 5 plugin instrumenting the profile/app
loader to do the needed replacement, potentially from its own service
loader (service dependent outside of karaf 5 structure).
Thinking out loud, it is very very close to tomee architecture which has
exactly that except the bootstrap loader(s) leaks a lot since it is assumed
not conflicting much and module services are not mainstream app oriented
but EE oriented. But the service and tree of loaders is there so overall if
Karaf5 would leverage the same architecture but power it by the Karaf core
feature inherited from OSGi which is to make any app running in the same
JVM.
A service example could be for example an instrumentation one (since
subclassloaders are karaf controlled) and do the javax -> jakarta migration
on the fly.


>
> regards
> Grzegorz Grzybek
>
>
> >
> > Romain Manni-Bucau
> > @rmannibucau  |  Blog
> >