On 09/11/2018 18:32, Brian Stansberry wrote:


On Fri, Nov 9, 2018 at 2:30 AM, Jean-Francois Denise <jdenise@redhat.com> wrote:
Hi,

in the galleon project, in a context where we would not provision a
complete server but just a subset required to run a given configuration,
we have identified a need for subsystems to advertise to the galleon
tooling some modules in addition to the modules galleon discover by
traversing the module dependency tree.

The first case is DeploymentProcessor injecting modules into the
deployment units (implicit modules).

The deployment injected ones are not required to be a dependency of the
subsystem module, so galleon has the risk to miss some of them.

As an example, in logging subsystem we have the following non optional
injected modules :
"org.jboss.logging",
"org.apache.commons.logging",
"org.apache.log4j",
"org.slf4j",
"org.jboss.logging.jul-to-slf4j-stub"

Some of these modules are direct dependencies of logging subsystem, some
others are indirect dependencies, others could be not present at all in
the module dependency tree (eg: org.jboss.logging.jul-to-slf4j-stub).

So we are thinking at solving this issue in 2 possible ways:

1) Mandate that all injected modules become dependencies of the
subsystem module at the cost to load some useless modules at runtime.

2) Possibly better, make the subsystem to call
RuntimeCapability.addAdditionalRequiredPackages (package name being
module name) for each injected module. An existing capability or a new
one  would have to be created.

There is also the case of optional injected module dependencies (eg: ee
subsystem org.glassfish.javax.el, org.eclipse.yasson, ...). We need to
treat them differently. When one is provisioning a server using galleon
he can choose to exclude some packages from the provisioned server.
Optional packages can be excluded without making the provisioned state
invalid (as opposed to required package that can't be excluded). These
optional implicit dependencies are typical usage of this, they are not
required by the deployment unit to properly operate.

For these, we plan to use
RuntimeCapability.addAdditionalOptionalPackages. We can't make them
"optional" in JBoss module. When galleon provision a subset of the
server we don't include all optional dependencies.

This brings the case of provisioning of optional dependencies present in
JBOSS module.xml.

We have identified multiple kind of optional dependencies.

1) The optional dependencies that are referencing modules only in use
when a feature is present in the configuration of the subsystem (eg: jmx
subsystem optional dep on org.jboss.remoting-jmx due to
<remoting-connector/>, remoting subsystem optional dep on
io.undertow.core for <http-connector/>, elytron subsystem optional dep
on org.picketbox for <jacc-policy>)

In order to have these dependencies to be provisioned with the
subsystem, we can attach thanks to
RuntimeCapability.addAdditionalRequiredPackages the modules to the
feature. When the feature is present in the configuration, the module is
no more optional but required.

2) The optional dependencies that reference modules that are part of
another subsystems and only use if this other subsystem is present (eg:
org.jboss.as.jpa optional dep on org.jboss.as.ejb3,
org.jboss.as.transactions optional dep on org.jboss.remoting). These
ones are simply not taken into account

3)  The optional dependencies that reference modules that are not part
of another subsystem (so are not provisioned by another source) but are
only meaningful if the other subsystem is present. We call these ones
"passive" (eg: org.jboss.as.weld optional dependency on
org.jboss.as.weld.ejb|jpa|beanvalidation|webservices|transactions).
"passive" are analyzed. If all their dependencies are present, then they
are included. Some implicit optional dependencies can fall into this
category (eg: org.jboss.jaxrs optional dep on
org.jboss.resteasy.resteasy-validator-provider-11).

The passive optional dependencies would be advertised with
RuntimeCapability.addAdditionalPassiveOptionalPackages(optional
dependency package name).

So to summarize:

- RuntimeCapability.addAdditionalRequiredPackages for all required
implicit modules

- RuntimeCapability.addAdditionalRequiredPackages to associate optional
dependencies to a feature

- RuntimeCapability.addAdditionalOptionalPackages for all optional
implicit dependencies that are not passive

- RuntimeCapability.addAdditionalPassiveOptionalPackages for all
optional dependencies (implicit or not) that are passive

A good question Alexey raised is whether the capability is the right concept for encapsulating this information.

In the end the key point is that the information ends up in the galleon feature-spec. That's the thing that encapsulates the configuration elements and the packages (i.e. fs content) that are needed to provide a feature. It's the feature that needs the package. We generate the galleon feature specs from our management model information. When in WF 13 we hit the first must-implement use case for getting additional package info into a feature spec I kind of arbitrarily picked 'RuntimeCapability' as the place to record that. It was a reasonable enough choice for that one use case but isn't great otherwise.

The alternative is to record this data in the ResourceDefinition, which is the primary source of the data that's used in the feature spec generation.

The conceptual argument is a capability is all about a named provider of a contract that other elements of the server can rely on; e.g. some MSC Service with a particular value type that will be exposed.  But this additional package stuff isn't really about that contract, it's about other things that will be done that likely are not relevant to other elements of the system.

A hint of a design smell would be if we end up creating RuntimeCapability instances for no other reason than to record this additional package information. Something I could see happening with this logging case you mentioned at the start.

I don't think ResourceDefinition has this problem.
I have introduced artificial capabilities in ee, jaxrs and weld (subsystems that are targeted by the demo) to add additional packages. It seems that it would be common case. So it is actually smelly.



Is it something that subsystems owner would be ready to put in place to
help galleon in this task? It would require a bit of analysis of the
dependencies of your modules but the gain could be quite important.

Please don't let my point above derail consideration of this question.  Use of ResourceDefinition vs RuntimeCapability is a relatively small detail. 

Some
early experimentation of this has shown a big reduction of the server
filesystem footprint (web server + cdi has been reduced from 156MB to
46MB). The runtime memory usage reduction is not that big, but less
modules being loaded we have a gain.

Thanks for reading.

JF

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



--
Brian Stansberry
Manager, Senior Principal Software Engineer
Red Hat