Regards JB
On 09/11/2015 09:26 AM, Christian Schneider wrote:
Hi Achim, very good overview of the problems users face and I also agree with your proposed solutions. Just for pax cdi I think we are not yet there so I would go with blueprint generation from CDI annotations for now and switch to proper pax cdi when it is fully working. On top I would add that I have the feeling that springsource more or less hyped microservices because they know spring does not scale! The bigger a spring application grows the more you see that it does not provide real modularization. 1. All of the dependencies have to be hammered into a flat classpath. This works or an astonishingly long time but at some point in the growth of a project it gets harder and harder. 2. The single spring context. In the end the whole application runs on one merged spring contex. All the settings done in the small contexts scattered around the project artifacts have to work together in the big context created at runtime. 3. No enforcement of visibility rules. You can define an API but you also have access to all the impls and developers tend to use them when the time pressure comes. 4. Static setup. Spring has to boot up all beans in the correct order. 5. They tried with OSGi but I think they realized that making spring fully OSGi ready is too much work All this leads to people thinking that an application needs to be split into microservices to make the above problems smaller. Only later they find that implementing microservices is pretty hard. The biggest problem is that they may not share a database. So you have to invent a lot of additonal logic to keep the different databases in the same state. Additionally there is security and monitoring which is also pretty much unsolved in microservices. If you compare this to a proper OSGi application then the need for the spring style microservices is almost not there. Instead you write small modules with a clearly defined API and boundaries and OSGi services to communicate. If you need to cross server boundaries then OSGi remote services allow to extend the OSGi services model to a larger environment. So I would argue that the microservices hype will not live to its promises. People will realize that OSGi provides better solutions with less overhead than microservices. This will take some time of course and we will need to make sure to not loose too many people to spring boot in the mean time. So what we need is a nice way to create microservices in karaf like you described. Additionally we should also allow people to just deploy a normal OSGi application in a simple way like you would for a microservice. So I think we need to make sure that karaf boot not only works for single projects but also for whole applications in form of a set of bundles. This will allow people to transition away from pure microservices when they realize they are not as good as they think while still keeping a lot of the benefits of the small and simple deployment unit. Christian Am 10.09.2015 um 23:58 schrieb Achim Nierbeck:Hi after reading this thread about two times, I finally understood the intention and the confusion in the thread. Let me try to summarize this to start some more confusion :-D Initial Problem: - learning OSGi has a steep learning curve - initial setup for a new OSGi project needs a lot of infrastructural "thingis" - doing the "microservice" dance is hard as it requires to much "extras" to have a simple REST service -- including infrastructure - Spring boot brings everything needed in one place in one "application" with just annotations These are the "problems" we know of and if I did get it right are the intention for the Karaf-Boot project. Now let me try to summarize the idea on how to solve that. From a simple Class file with some annotation generate - a OSGi service, including OSGi wiring via Blueprint - via an extra maven plugin - get rid of all those infrastructural "add-ons" for the first OSGi bundle by having a easy to use Parent-POM - do the microservice - dance by wrapping everything together including CXF in a "one-size-fits-all" executable jar, which is a wonna-be-spring-boot jar but made of karaf Now some thoughts of mine ... I somehow fear you try to mix several problems and try to solve all of them with one go. Let's try to analyze those initial problems separately and we might find a valid approach. Learning curve of OSGi: right now we have different approaches to make it easier with OSGi. - Blueprint - DS - CDI - iPojo - Felix DependencyManager Blueprint just tastes to much like spring xml and actually isn't state of the art anymore. You'll get a lot of negative feedback regarding the "old-school" xml. It's biggest plus is the Enterprise "features" like JPA and JTA, which isn't covered by any other technique right now (unless you look at non Apache Hosted OSS projects like Amdatu). Declarative Services, especially the latest R6 looks rather promising, especially in conjunction with MetaType. But we have specialized annotations which again don't actually help with the learning curve. Another downside is the missing JPA and JTA Enterprise features. CDI, we have the Pax-CDI project which has been on the way of being the reference implementation of the OSGi-CDI spec. Well, we all know how that turned out. One of its biggest benefits is definitely the easy transition because of the CDI annotation which help a lot with wiring inside a bundle and also with extra annotations (@OsgiService) with OSGi-Services. CDI though still lacks runtime OSGi annotation support for JPA and JTA if we are somehow capable to compensate that, we'll have the required EE stack at hand. iPojo ... sorry never used Dependency Manager is a bit more low-level compared to DS but seems to have a nice DSL for easy wiring of Services. So do we really need another Build-Time-Annotation which generates another "wiring" plan which is used during runtime (exactly what DS already does)? Right now I'm not convinced of this. Instead of another generation I would prefer to use CDI as it is known to work in classic Spring/EE world and OSGi. Instead we should focus on bringing the full support of Enterprise style annotations to the OSGi world. Initial Setup for a new OSGi Bundle Project Instead of a parent POM I think it would be much more beneficial to work with a BOM (Bill of Material) for easier development, especially when working with std. CDI annotations plus some extra PAX-CDI ones. Maven Archetype are actually good enough to create an initial structure for a maven project. The big benefit of the spring-boot approach is to actually have a bunch of different BOMs that can be used to add certain functionalities to the project, aka spring-boot-*-starter. To start a karaf-boot I think we should first try to have a karaf:run available as maven plugin. The microservice dance actually it's just add a rest-service on top of a OSGi service, that's all that is needed in Karaf. Right now I'm not in favor of any certain framework. CXF seems a bit bloated but is known to work, but requires blueprint. Jersey, I've seen that to be working with "plain" OSGi. A bit of polishing and it should be quite easy to use, especially with CDI at hand. But it needs more to dance the microservice dance, you need "small" containers ... which is quite contrary to the way Karaf and OSGi in general is working with services. But this is the point I think the karaf profiles come in handy. You don't need a full blown Karaf, just a basic infrastructure with your own Bundle, might as well ignore the shell. In the end dump that into a docker container and if you need to do a bugfix do it the "docker" - way. spring-boot brings it all in one go karaf-boot should do the same, but actually I fear we do more then needed. For a new Project setup I'd rather would like to see different karaf-starter-* BOMs and a karaf:run maven plugin Some more docuementation for the profiles of Karaf could also be helpful :D to build minimalistic karaf instances runnable in docker containers. Regarding the karaf:run it might be interesting to "re-activate" the pax:run maven plugin to run nicely with a karaf instance, or use it as foundation for the karaf:run maven plugin. So in the end, do we really need all this? I'm not so sure, but we surely need an easier "to use" approach. Therefore we should first focus on having easier setup of bundle development. -> karaf-boot-starter-* BOMs should take care of that -> karaf:run should make it easier to have a running container Do we need new annotations? I can't see that yet. Instead we should/could focus on the following: a) make sure DS also is capable to work with JPA/JTA and some other enterprise annotations b) make sure CDI works with runtime JPA/JTA annotation smoothly c) provide more demos and archetypes with OSGi and CDI annotations regards, Achim 2015-09-10 20:41 GMT+02:00 Jean-Baptiste Onofré <[email protected]>:Thanks Milen, it's an open place to discussion. We just share standpoints and opinions: that's the key part !!! My goal is to give more traction on Karaf by providing easier We all provide valid points, but I think we are not the most setted to argue as we are deeply involved in OSGi and Karaf. The karaf-boot proto came when discussing with new Karaf users, coming from "outside" of OSGi, or people that decided to use spring-boot (even if they like a lot Karaf), just because it's easier. It's hardly frustrating for us as we just need some tooling to provide even more traction. On the container side, I think Karaf is already great, and answers all needs. The part where we should improve what we deliver is around developer tooling: easier, faster, key turn. If karaf-boot will be a success, I don't know (who knows ? ;)). But anyway, it brings points, questions, and identify some missings in the current picture. My $0.01 ;) Regards JB On 09/10/2015 08:02 PM, Milen Dyankov wrote:Well I was just referring to your example but I get your point. Which reminds me of EnRoute <http://enroute.osgi.org/> project which despite the big names and the most popular OSGI build tool behind it, doesn't seem to get as much traction as I expected! That said, I really admire your enthusiasm and wish KarafBoot can be more successful that that. I'm not trying to discourage you! Just it seams what you are after is something that other people have tried already with questionable success. Best, Milen On Thu, Sep 10, 2015 at 7:22 PM, Jean-Baptiste Onofré <[email protected]> wrote: And how to you deal with jpa, jta, rest, etc with SCR annotations ?Regards JB On 09/10/2015 07:16 PM, Milen Dyankov wrote: So correct me if I'm wrong but if I get the sample you provided in thefirst mail and replace: - the parent pom with "maven-bundle-plugin" - @Bean with @Component - @Init with @Activate wouldn't that have the exact same end result? I mean it obviously differ in terms of what gets generated (Blueprint vs DS) but form end user perspective there is no difference, right? On Thu, Sep 10, 2015 at 6:55 PM, Jean-Baptiste Onofré <[email protected]> wrote: Hey Milen,Actually, there's too part: 1/ karaf-boot-starter will do the ready to start artifact, embedding karaf, but it's another point 2/ the value of karaf-boot annotations and plugin is first to simplify the bundle/artifact ready to be deploy-able into Karaf (generate the "plumbing" easily for developers). Regards JB On 09/10/2015 06:50 PM, Milen Dyankov wrote: " ... that you deploy in Karaf ..."OK may be I misunderstood the concept. I thought the result is standalone executable JAR, thus my comments above. If on the other hand I need to install Karaf and then deploy my services into it I really don't see how it differs form what people are doing now? I'm sorry if I'm not making much sense. I didn't have the time to experiment with your code and samples so may be I'm missing an important peace here. Best, Milen On Thu, Sep 10, 2015 at 6:27 PM, Jean-Baptiste Onofré < [email protected]> wrote: Allow me to disagree: Karaf is a perfect container for microservices.Image to create a microservice (using karaf-boot) that you deploy in Karaf and use such service in another microservice, all wired with OSGi service and Karaf: we leverage OSGi/Karaf as a microservices container. But even without talking of microservices, new developers to Karaf (and OSGi generally speaking) are frustrated by the effort on non business code to do (I have to write an Activator, or a descriptor, etc, etc). So, a tooling to simplify this is still a valid addition IMHO. Regards JB On 09/10/2015 06:23 PM, Milen Dyankov wrote: I might be wrong but I think the whole success of SpringBoot (apart from having the "Spring" in it) is the microservices hype!it's quick and easy but most usecases follow the "create one (or very few) service(s), pack them as single executable and access them via REST" pattern. We can obviously do the same with OSGi and Karaf in particular but personally I think this makes absolutely no sense. In such approach one in not benefiting form OSGi almost at all. Honestly speaking I would argue that if one does not understand how OSGi service layer works (regardless of the framework used to register/access services) it makes no sense to use OSGi at all. Just my 2 cents! Regards, Milen On Thu, Sep 10, 2015 at 6:08 PM, Christian Schneider < [email protected]> wrote: I already created such a maven plugin in aries. The user can use standard CDI and JEE annotations and the result is blueprint xml.How is the new approach different / better? Why should it be good for the developer to move away from well defined standard annotations and use custom annotations that bind him to karaf? I mean if this is created by the spring guys I know they want to catch people by perceived simplicity and then make sure to make it difficult to switch. As an open source comminity I do not know why we should do something like this. Abstracting away from frameworks just means you create another layer that people then also have to learn. There were some cases in the past where this make sense because the underlying frameworks sucked (like JEE 2). This is not the case today though I think. What kind of use case do you have in mind? Every project starts small but it needs to be able to grow then. You can not start with custom annoations and then tell people to later switch to something else when the project grows. I think it makes more sense to make it easier for people to use the standard annoations and use the right dependencies. If we simply provide a tooling that makes it easy to start with SCR or blueprint we provide much more value for people as thery can then grow without any breaking changes. Christian Am 10.09.2015 um 17:46 schrieb Jean-Baptiste Onofré: Because all these annotations are runtime: here we talk about tooling at build time.More over, the purpose is to provide more high level annotations, which abstract actual annotations/frameworks that we can use under hood. The purpose of centralizing all in karaf-boot is to have a central project: the developer just use karaf-boot, it doesn't really know what technologies are involved behind the scene. For instance, in spring-boot, they use activemq, jersey, etc, but all from spring-boot. The developers don't know a rest service use jersey for instance, it's completely abstracted. Again the purpose is to simplify life for developers: splitting the annotations in different projects introduces complexity (at least to find the dependencies and core import packages). If an advanced developer wants to use CDI, SCR, etc, he can of course. Regards JB On 09/10/2015 05:40 PM, Christian Schneider wrote: I am not really enthusiastic about duplicating functionality of cxf or aries. Aries supports a very nice approach for injections, jpa andjta. Why should it make sense to recreate that? Aries blueprint also has annoation support even in two flavors (CDI, custom). How does the new approach interact with this? Instead I propose we create support for such annotations in the respective projects (where they are missing) and concentrate on karaf as a container not an application development framework. By leveraging the existing frameworks we profit from their own development teams. Whatever we recreate will have to be developed by the very few resources of the karaf team. Christian Am 10.09.2015 um 16:53 schrieb Jean-Baptiste Onofré: Hi Guillaume, thanks for your feedback.I fully agree about providing more high level annotations (it's what I do with @jpa, @rest, @soap, @jta annotations). I agree that the current annotations are too low level, and blueprint "oriented". I just move forward a bit with the current codebase, just to illustrate karaf-boot usage in the samples. But again, you are right, and I will create a new annotations set. One of the purpose of karaf-boot annotations is to "abstract" the actual code/artifact that we generate. So, if now we generate blueprint, without changing the karaf-boot annotations, we will be able to generate something else (why not SCR, etc). I agree with a BOM, but I think it's interesting to provide both: - providing a ready to use parent pom allows developers to create a very simple pom.xml where all plugins and dependencies are already defined - for more advanced devs, they can create their own pom.xml starting from the BOM or archetype. Thanks again for your feedback ! Regards JB On 09/10/2015 04:44 PM, Guillaume Nodet wrote: I like the idea. For the annotations, we need to keep really high level. Theannotations in the code base right now are much too close to blueprint. I think we need to grab a small enough subset so that the annotations are easy to understand for beginners and without any ambiguities, even at the cost of features. For example, I think we should restrict to constructor injection, so that we don't have any bind / rebind / init methods. We simply need an optional @Destroy. In case the dependencies change at runtime, simply destroy the bean / service and recreate it the dependencies are still met after the change. If blueprint is to be hidden completely, we may find a better alternative in SCR or even Felix Dependency Manager, but it does not matter too much for now. I agree with the idea of using a BOM instead of a parent if possible. I'm not very familiar, but this is less invasive. The real problems will come with the support of higher level annotations for JAXRS, JPA, etc... Not really sure how to handle those yet... 2015-09-09 16:32 GMT+02:00 Jean-Baptiste Onofré < [email protected]:Hi all, I worked on a prototype about Karaf Boot.Let me give you some backgrounds and discuss about that all together. Why Karaf Boot ? ---------------- When you develop artifacts (bundles) to be deployed in Karaf, you can see that the actual time that you spend on your business code is finally largely less important that all the plumbing effort that you have to do (writing OSGi Activator, or blueprint/scr descriptor, etc). It means that your "go to market" is longer, and we should provide something that allows you to focus on your code. Even if SCR annotations is a very good step forward, some use cases are not so easy to do (JPA, JTA for instance). And anyway, you have to prepare your pom.xml with different plugin and dependency. Moreover, when you have your artifacts, you have to prepare Karaf container, and deploy those artifacts there. Even if it's "container" approach is the most important for me, we can give even more flexibility by providing a way to embed and prepare Karaf in a ready to execute jar/artifact. What is Karaf Boot ? -------------------- Karaf Boot provides four components: * karaf-boot-parent is the Maven parent pom that your project just inherit: that's all ! All plugins, dependencies, etc are described in this parent, you even don't have to define packaging as bundle, standard jar is fine. * karaf-boot (coming with karaf-boot-parent) provides annotations that you use directly in your business code (like @Bean, @Service, @Reference, @Inject, etc): again, your focus on your code, karaf-boot deals with the plumbing. * karaf-boot-maven-plugin (coming with karaf-boot-parent) scan the classes and generate a blueprint XML. For now, I'm using blueprint generation (because we can cover lot of use cases, for instance, I plan to provide @rest annotation that will generate blueprint XML with cxf jaxrs server, etc). * karaf-boot-starter is the module providing a convenient way to embed, configure and bootstrap Karaf. Just to illustrate this, let's take a look on the karaf-boot-sample-simple. The pom.xml is really simple: <?xml version="1.0" encoding="UTF-8"?> <project xmlns="http://maven.apache.org/POM/4.0.0"; xmlns:xsi=" http://www.w3.org/2001/XMLSchema-instance"; xsi:schemaLocation=" http://maven.apache.org/POM/4.0.0 http://maven.apache.org/xsd/maven-4.0.0.xsd";> <modelVersion>4.0.0</modelVersion> <parent> <groupId>org.apache.karaf.boot</groupId> <artifactId>karaf-boot-parent</artifactId> <version>1.0.0-SNAPSHOT</version> </parent> <artifactId>karaf-boot-sample-simple</artifactId> <version>1.0.0-SNAPSHOT</version> </project> You can see, the only thing that the developer has to do: define karaf-boot-parent as parent pom. That's all. Now, in the code, you have just one bean that we want to run: package org.apache.karaf.boot.sample.simple; import org.apache.karaf.boot.Bean; import org.apache.karaf.boot.Init; @Bean(id = "simple-bean") public class SimpleBean { @Init public void simple() { System.out.println("Hello world"); } } You can see the @Bean and @Init karaf-boot annotations. The karaf-boot-maven-plugin will generate the blueprint descriptor using this. Current Status -------------- I pushed Karaf Boot structure there: https://github.com/jbonofre/karaf-boot It's a mix of rewrapping of existing code (from aries, pax-exam, etc) and additions. I created the annotations, I'm now working on the karaf-boot-maven-plugin based on Christian's work in aries (I'm actually scanning the boot annotations now, and generating the XML). I will push new changes later today and tomorrow. Open Questions --------------- * For now, I would prefer to be 'artifacts' and 'resources' generator: I think it's better than to depend to a feature running in Karaf, but it's open to discussion. * I'm now generating blueprint. Probably native OSGi or scr generation can make sense. * I'm generating bundles: thanks to the Karaf4 features resolver, as the bundles provide requirements/capabilities metadata, I think it's a good start. However, maybe it's worth to be able to create features, kar, profile. Thoughts ? Thanks, Regards JB -- Jean-Baptiste Onofré [email protected] http://blog.nanthrax.net Talend - http://www.talend.com--Jean-Baptiste Onofré[email protected] http://blog.nanthrax.net Talend - http://www.talend.com--Jean-Baptiste Onofré [email protected] http://blog.nanthrax.net Talend - http://www.talend.com--Jean-Baptiste Onofré [email protected] http://blog.nanthrax.net Talend - http://www.talend.com-- Jean-Baptiste Onofré [email protected] http://blog.nanthrax.net Talend - http://www.talend.com
-- Jean-Baptiste Onofré [email protected] http://blog.nanthrax.net Talend - http://www.talend.com
