I think it's the right direction. It will just take some care to not
break stuff by not registering stuff that really should be registered.

Good news is I had the same concern about
https://issues.jboss.org/browse/WFCORE-831 but it seems there wasn't
much of a problem.

Twists to think about in this:

1) The "runtime-only but ok on the DC" case I mentioned, but I think
that's solvable with more metadata, and we don't support it now so you
won't break anything.

2) Subsystems running on the HC. I don't have any good reason why these
would be an issue; I'm just pointing them out as something to remember
as pretty much no one but Kabir has dealt with them. The
context.isRuntimeOnlyRegistrationValid() provides the correct response
for those.

3) Definitions of runtime-only child resources. This one is a bit more
tricky. To be consistent they should be handled similarly to how you
propose dealing with metrics. But they are different -- the code that
registers the runtime-only ResourceDef may then go on and try to use the
returned MRR, or may even come back later and look it up. But really the
MRR shouldn't end up in the tree at all.

Good news about 3) is code that is doing what I describe is probably
already guarding that work with
context.isRuntimeOnlyRegistrationValid(). So changing something probably
won't break that code.

4) API change. Doing this is going to introduce management API changes,
by removing stuff that should have never been there. So we need to
account for that (probably bump API versions, check for console breakage.)


BTW, the datasource subsystem thing is a bit of a different beast. The
actual list of metrics is not known until runtime, because they vary
based on the driver. That's why the registration is done that way.

- Brian

On 4/25/16 10:15 AM, Jeff Mesnil wrote:
> TL;DR - I propose to simplify subsystem development by moving some of the validation logic from the resource definitions to the management resource registration. The goal is to provide a static representation of the resources and let the MMR dynamically pick the “meaningful” parts.
>
> Last week an user complains that the messaging-activemq subsystem’s statistics were not updated in domain mode.
> It turned out that he was reading the metrics on the DC (/profile=full/subsytem=messaging-activemq/…) instead of reading on the server (/host=master/server=server-one/subsystem=messaging-activemq/…)
>
> It is a bug[1] in the messaging-activemq because its resources register their metrics without checking whether that makes sense. The correct check is to look at context.isRuntimeOnlyRegistrationValid() to check whether a resource can register its metrics (the same check applies also for runtime attributes/operations).
> I looked at other subsystems and undertow has the same issue.
>
> This check does not work well with the PersistentResourceDefinition that is used throughout the messaging-activemq and undertow subsystems. This API works best when the definition of the resources uses a bunch of static instances for the resource itself, its attributes, metrics, etc. These static instances are also used by the companion PersistentResourceXMLDescription to provide a static XML representation of the subsystem.
> If I have to pass this context.isRuntimeOnlyRegistrationValid() boolean to every resources in the subsystem, I get rid of the static representations used by the PersistentResourceDefinition and PersistentResourceXMLDescription and I have to add a lot of simple but boilerplate code in all my resource definitions.
>
> The datasources subsystem does not exhibit this issue. It works around it by installing a Service at RUNTIME step to register (resp. unregister) statistics resource definitions when the service is started (res. stopped). Since services are only installed when runtime is supported, it ensures that the datasources metrics are available only on server and not on the DC.
> It looks correct but I’m not a big fan of this solution. It makes the subsystem definition more complex to understand and it also involves boilerplate code that every subsystem providing runtime should write.
>
> I was wondering if we could simplify the development of the subsystems by moving some of the logic dealing with that use case in the ManagementResourceRegistration instead.
>
> My proposal would be to have the resource definitions be “complete”. The resource always defines its attributes/metrics/operations.
> When the MMR takes this definition and registers the different parts, it would only register the “meaningful” one depending on the ProcessType and RunningMode. E.g. the MRR of the DC or a admin-only server would not register metrics & runtime attributes/operations while the MMR on a normal server would register everything.
> This increase the complexity of the MMR which has to figure out whether to actually register something or discard it but it makes writing subsystem simpler and more robust.
>
> Brian told me there might some exceptions (e.g. a runtime op that could be invoked on the DC) but these case could be handled by adding metadata to the resource definitions.
>
> This approach segues quite well with the idea to generate subsystem using annotations. All the subsystem developers has to do is describe extensively the subsystem resources (using static objects now, annotations in a future version) and let the MMR decides which parts of the resources are actually registered.
>
> To sum up, the definition of a resource is static, while its registration is dynamic.
>
> Do you see any issue with that proposal?
> As a first step, I’ll start by creating the corresponding WFCORE issue to fix WFLY-6546 issue by ensuring the MMR does not register metric if the runtime registration is not valid. This should not have any API impact (but the behaviour will certainly change).
>
> jeff
>
> [1] https://issues.jboss.org/browse/WFLY-6546
>


--
Brian Stansberry
Senior Principal Software Engineer
JBoss by Red Hat

_______________________________________________
wildfly-dev mailing list
wildfly-dev@lists.jboss.org
https://lists.jboss.org/mailman/listinfo/wildfly-dev