----- Original Message -----
From: "Alon Bar-Lev" <alonbl@redhat.com> To: "Vojtech Szocs" <vszocs@redhat.com> Cc: devel@ovirt.org, "Mark Proctor" <mdproctor@gmail.com> Sent: Wednesday, November 5, 2014 5:16:15 PM Subject: Re: [ovirt-devel] Thoughts on modularization
----- Original Message -----
From: "Vojtech Szocs" <vszocs@redhat.com> To: "Alon Bar-Lev" <alonbl@redhat.com> Cc: devel@ovirt.org, "Mark Proctor" <mdproctor@gmail.com> Sent: Wednesday, November 5, 2014 6:07:50 PM Subject: Re: [ovirt-devel] Thoughts on modularization
----- Original Message -----
From: "Alon Bar-Lev" <alonbl@redhat.com> To: "Vojtech Szocs" <vszocs@redhat.com> Cc: devel@ovirt.org, "Mark Proctor" <mdproctor@gmail.com> Sent: Wednesday, November 5, 2014 4:32:31 PM Subject: Re: [ovirt-devel] Thoughts on modularization
----- Original Message -----
From: "Vojtech Szocs" <vszocs@redhat.com> To: "Alon Bar-Lev" <alonbl@redhat.com> Cc: devel@ovirt.org, "Mark Proctor" <mdproctor@gmail.com> Sent: Wednesday, November 5, 2014 5:24:14 PM Subject: Re: [ovirt-devel] Thoughts on modularization
----- Original Message -----
From: "Alon Bar-Lev" <alonbl@redhat.com> To: "Vojtech Szocs" <vszocs@redhat.com> Cc: devel@ovirt.org, "Mark Proctor" <mdproctor@gmail.com> Sent: Wednesday, November 5, 2014 4:12:06 PM Subject: Re: [ovirt-devel] Thoughts on modularization
----- Original Message -----
From: "Vojtech Szocs" <vszocs@redhat.com> To: devel@ovirt.org Cc: "Mark Proctor" <mdproctor@gmail.com> Sent: Wednesday, November 5, 2014 5:04:24 PM Subject: [ovirt-devel] Thoughts on modularization
Hi guys,
I've discussed this recently with Yair and Mark, I just wanted to share some more thoughts on this topic -- in particular, how modularization problem can be approached (regardless of implementation details).
I see two approaches here. The typical one is to define APIs for modules to consume. For example, oVirt Engine extension API has API for auth stuff; oVirt UI plugin API has API for showing tabs and dialogs, etc. The advantage is strict consistency, disadvantage is burden of having to maintain the whole API. With this approach, you tell modules: "This is the API to work with system, defining how you can plug into it."
Now turn 180 degrees. The other approach, which is really interesting, is to let modules themselves export API. This naturally leads to module hierarchies. Ultimately, this leads to micro-kernel-style development, where all logic resides in modules. Now you might ask: "What if we want to employ some consistent work flow across multiple modules? For example, have some pluggable *auth* infra?" -- this can be done via some "higher" level module, that exports API and "lower" level modules consume that API.
If you have any ideas, please share!
Both solutions can be applied using existing extension api, an extension can locate other extension and interact with it the same way the core interacts with extensions.
But how does core interact with extensions? I assume via well-defined API, i.e. in accordance with first approach mentioned above.
presentation: http://www.ovirt.org/File:Ovirt_3.5_-_aaa.pdf
Thanks for sharing!
package org.ovirt.engine.api.extensions;
/** * Interface of an extension. */ public interface Extension {
/** * Invoke operation. * @param input input parameters. * @param output output parameters. * * <p> * Interaction is done via the parameters. * Exceptions are not allowed. * </p> * <p> * Basic mappings available at {@link Base}. * </p> * * @see Base */ void invoke(ExtMap input, ExtMap output);
}
OK, so it was my lack of knowledge of AAA implementation :)
Since Extension interface is invoke-based, I assume that Engine core (backend) invokes extensions on specific occasions, right?
correct, and nothing prevents extension to invoke other extensions.
I see. I just wanted to clarify who makes the initial invocation. If there is some core, extensions just extend that core. This is different from my view of pure modularization, where there is no core, there are just modules and all important logic resides in modules.
With second approach mentioned above, core would not interact with extensions at all (or in a very limited way), instead - extensions would interact with each other. In other words, extension would not need to implement core-specific API (there would be none), instead it would inject its dependencies (other modules/extensions) and consume their APIs. This is the difference I wanted to point out :)
The extension interface is primitive to enable exactly that, provided java people will open their minds :)
Simple interface means great flexibility, now I understand the rationale behind typed maps in extension API.
What I had in mind was something like this (maybe crazy):
* imagine for a moment that the backend is fully modular (logic lives in modules, no logic outside modules)
* @DependsOn("SomeOtherModule") // modules can form hierarchies public class MyModule implements Module { public void loaded() {} // executed just once per module }
you again using java magic for something that can go to declaration...
Java annotations are static metadata, they are not magic :) Annotations are helpful because they declare what should be done, not how it should be done. @DependsOn in example above means that some infra will process/scan classes and if it finds this annotation, it will invoke extra logic according to the semantics of this annotation. Yes, it could also go into "conventional" declaration like properties files etc. I have nothing against that, the major point is declaring useful metadata in some way :)
* API for fetching modules, i.e. getModuleByName('xxx')
there is no need as within the extension context you have access to this information, all you need is iterate a list, see Base.java.
What I meant: module A exports API "foo", module B depends on module A and therefore can consume API "foo". IIUC, in Engine ext-api you must find appropriate extension yourself, in order to invoke it. This is the difference, because what I suggested above is to allow modules to state their dependencies, instead of having to locate them on their own. Declaring dependencies in advance has many advantages, one of them is you can avoid "always load all dependencies in some sequence" because you don't know which module will need which other module.
API for exporting module API, i.e. exportApiObject(foo)
not sure why it is needed... once extension is loaded it is available.
It was meant as means of communication between modules. Engine ext-api invoke-based interface defines communication between extensions via input & output maps. What I meant is different mean of communication -> one module exports its API (i.e. object) and another module can consume that API.
This way, you'd have one "core" module, providing most general functionality/abstractions. Then you'd have more specific modules, plugged into "core" module's API etc. In other words, hierarchy of modules. This is essentially modularization taken to extreme :)
(BTW, above is just an idea.)
not sure I understand what is the functionality difference between this and the primitive approach we already have...
The difference is: a, have some (monolithic) core, which invokes extensions b, have no core, everything is a module
1. every configuration at /etc/ovirt-engine/extensions.d and /usr/share/ovirt-engine/extensions.d are loaded during engine startup.
In a modular architecture, where everything is a module, module "tree" is determined (according to module dependency declarations) and loaded. In a typical extension architecture, all extensions are always loaded in some sequence.
2. each extension has two phases load and init, during init an extension can locate other extensions by name via its context. it is true that probably missing functionality is to support ordering of the init and load, we left it into future as there was no actual need for this.
This is actually quite similar to how our UI plugins are loaded, except for that "locating other plugins" is not implemented.
3. each extension can interact with other extension.
4. there must be core model to trigger the entire thing, core cannot be just a loader.
This is a design decision. A "core" can be either monolithic base (includes business logic) that invokes extensions. Alternatively, "core" can be just module loader, and all logic is in modules. So I disagree - core CAN be "just" a module loader. See http://requirejs.org/docs/whyamd.html for details on this topic.
Alon