The work to configure and get the module only needs to be done ones - a
module will never "unresolve" unless you undeploy it.
11:18:05 PM)*tdiesler:* Nihility, for osgi the behavior for a bundle
deployment that cannot resolve would be such that the deployment
does not fail but it hangs in some pre_module phase. you could
deploy additional stuff and retry to resolve/start the thing through
the console. It is metadata valid and stays in the runtime as such.
(11:18:54 PM)*tdiesler:* That would solve all ordering issues I think
(11:19:42 PM)*Nihility:* yeah so like if we had a dep on another
deployments MODULE
(11:20:03 PM)*tdiesler:* you dont get an error
(11:20:08 PM)*Nihility:* that would do what you are talking about
(11:20:13 PM)*tdiesler:* yes
(11:20:13 PM)*Nihility:* it just waits
(11:20:29 PM)*Nihility:* of course we have a circularity problem to
think about
(11:20:37 PM)*Nihility:* I'm not sure if we support this today or not
(11:20:58 PM)*Nihility:* but i wonder if a.jar (dependency b) b.jar
(dependency a) works for modules
(11:21:04 PM)*Nihility:* we make it work for class-path
(11:21:15 PM)*Nihility:* since Class-path is supposed to allow it
(11:21:32 PM)*Nihility:* it works in jboss modules
(11:21:41 PM)*Nihility:* but as a deployment module dep I'm not sure
(11:21:51 PM)*Nihility:* stuartdouglas: would know, are you up?
(11:22:19 PM)*Nihility:* tdiesler: i take it thats completely
invalid with bundles right, circularity?
(11:22:39 PM)*tdiesler:* Nihility, it works in osgi. because the
resolver knows the complete set of metadata. and when triggered to
resolve a set of bundles and it finds a cosistent solution we go to
the jboss-modules PAI
(11:23:10 PM)*tdiesler:* Nihility, I don't think so
(11:23:24 PM)*Nihility:* ah ok
(11:23:37 PM)*tdiesler:* Require-Bundle is of course discouraged
anyway, but it should work
(11:23:58 PM)*Nihility:* tdiesler: you can do it with packages too
right?
(11:24:12 PM)*tdiesler:* yes
The default behaviour for a bundle deployment would be "don't fail on
cannot resolve". For other deployment types it could be different as it
is now. An explicit start op would take the deployment past the resolve
phase and through the activate phases. If it still cannot resolve you'd
get an error.
On the CLI you can do:
deploy foo.jar --no-start (which means it is ok that foo cannot
resolve)
deploy bar.jar
start foo.jar (which means that foo must resolve and start without error)
stop foo.jar (which means that foo runs back through the activate
DUPs only - it
stays resolved and keeps the module)
The cleanup issue only becomes a problem if you want to do repeated
start ops.
On 09/21/2012 01:59 AM, Stuart Douglas wrote:
I did no a similar thing a little while ago, that would allow a
deployment with a dependency on another deployment to basically sit
and wait for the next deployment to come up.
I think the way to implement this is to have the configure
CONFIGURE_MODULE phase be set to passive, so it only starts
when the services that it depends on become available. So in the OSGI
case this would depended on the RESOLVE services, and would start as
soon as resolution completed.
Things get more complicated if this passive service has its
dependencies go away, as deployment phases are not restartable. I
think the only way to actually handle this is to perform a full
re-deployment. The reason why this is so problematic is because the
DUP contains a lot of mutable state that is modified in each phase.
Re-running all the DUP's would result in a lot of this information
being doubled up.
It is also more complicated if we are dealing with a multi-module
deployment. If any of the sub deployments are not set to start lazily,
and are dependent on services that are present in the lazy deployment
then this will fail. This will probably not be a big deal in practice.
Another big issue is that we have no way of reporting the success or
failure of the deployment back to the user. We also need to make sure
that if this passive deployment fails we have some way of rolling back
the deployment, to keep the behavior consistent with normal
deployments. We do not want to leave a half finished deployment
sitting around.
Basically I think this is possible for a fairly limited case where:
- The deployment stops after configure module
- We have a mechanism in place to prevent the DU from attempting to
restart without a full re-deployment.
- We make sure that partial deployment is not allowed, and a full
rollback is attempted if start fails.
Stuart
Scott Marlow wrote:
> On 09/20/2012 02:52 PM, Thomas Diesler wrote:
>> On 09/20/2012 02:29 PM, Scott Marlow wrote:
>>> On 09/20/2012 09:18 AM, Thomas Diesler wrote:
>>>> Folks,
>>>>
>>>> I rebased the
>>>>
as2777<https://github.com/tdiesler/jboss-as/tree/as2777>
>>>> branch on the current as3694
>>>> <
https://github.com/tdiesler/jboss-as/tree/as3694> approach.
>>>>
>>>> This integrates with the CLI and you can do
>>>>
>>>> <deployment name="simple.war"
runtime-name="simple.war">
>>>> <content
>>>> sha1="be6f7ba1ceaa2028c86d080648e1467904999f8e"/>
>>>> <properties>
>>>> <property name="start.policy"
value="deferred"/>
>>>> </properties>
>>>> </deployment>
>>>>
>>>> [tdiesler@tdvaio jboss-as-7.2.0.Alpha1-SNAPSHOT]$ bin/jboss-cli.sh
>>>> --connect
>>>> [standalone@localhost:9999 /] start simple.war
>>>> [standalone@localhost:9999 /] stop simple.war
>>>>
>>>> on the server you should see
>>>>
>>>> 08:56:30,002 INFO [org.jboss.web] (MSC service thread 1-2)
>>>> JBAS018210:
>>>> Register web context: /simple
>>>> 08:57:16,533 INFO [org.jboss.web] (MSC service thread 1-3)
>>>> JBAS018224:
>>>> Unregister web context: /simple
>>>>
>>>> The current implementation sets the initial mode of the
>>>> CONFIGURE_MODULE
>>>> to NEVER for start.policy=deferred
>>>> The start/stop operations toggle the mode of the INSTALL phase
>>>> service
>>>> between ACTIVE/NEVER.
>>>>
>>>> The CLEANUP phase is currently a problem on repeated start. You'll
>>>> see
>>>>
>>>> Caused by: java.lang.NullPointerException
>>>>
atorg.jboss.as.ejb3.deployment.processors.merging.AbstractMergingProcessor.deploy(AbstractMergingProcessor.java:61)
>>>>
>>>>
>>>> at
>>>>
org.jboss.as.server.deployment.DeploymentUnitPhaseService.start(DeploymentUnitPhaseService.java:124)
>>>>
>>>>
>>>> [jboss-as-server-7.2.0.Alpha1-SNAPSHOT.jar:7.2
>>>>
>>>> because the EEModuleDescription is no longer attached.
>>>>
>>>> Stopping a deployment would not need to take the Module and
>>>> associated
>>>> ClassLoader down. So any datastructures needed to configure the
>>>> Module
>>>> can safely be cleaned up IMHO. I presume the DUPs can reasonably be
>>>> fixed by moving them to the appropriate phases (see below) and by
>>>> making
>>> Which DUPs do you propose moving and to which phase? Currently, the
>>> first DUP to use a deployment module classloader, is JPA (for class
>>> enhancing (rewriting) in the FIRST_MODULE_USE phase as part of
>>> creating the EntityManagerFactory). If class enhancing is not
>>> enabled, the JPA entity manager factory is created during the INSTALL
>>> phase.
>> All DUPs currently assume that they run in a single transaction. IOW,
>> the DUPs in the INSTALL phase are not designed for the deployment to
>> yo-yo in that phase, which they would have to do if we are going to
>> implement deployment start/stop behaviour (without reconstruction on
>> the
>> Module) properly.
>>
>> I suggest the review the DUPs in the INSTALL phase to see whether they
>> use data structures that are removed in the CLEANUP phase. There are
>> two
>> possibilities
>>
>> #1 DUP (in INSTALL) should be moved to a phase that is only executed
>> once. Then the associated data can safely be discarded in CLEANUP
>>
>> #2 DUP (in INSTALL) should only use data that stays attached with
>> the DU
>> and is not discarded in CLEANUP
>>
>> I guess this has to be decided on a case-by-case basis. Each subsystem
>> would have to make up its mind what it means to start/stop their
>> respective deployments. For a webapp that probably means to start/stop
>> the web context.
>
> JPA class enhancing/rewriting (if enabled), needs to happen before any
> application classes are loaded. The JPA class enhancing/rewriting
> occurs when the EntityManagerFactory is created.
>
> If we were to allow non-JPA deployers to run in the same or earlier
> phase as JPA, we would need a strong guard to prevent the non-JPA
> deployers from invoking any code that could load application classes
> (while letting JPA to load application classes).
>
>>>> sure they still have the data structures attached to do yo-yo on the
>>>> activate phases.
>>>>
>>>> Using this
terminology<https://issues.jboss.org/browse/AS7-3585>
>>>>
>>>> STRUCTURE => STRUCTURE
>>>>
>>>> PARSE => PRE_INSTALL
>>>> REGISTER => INSTALL
>>>> => POST_INSTALL
>>>>
>>>> DEPENDENCIES => PRE_RESOLVE
>>>> CONFIGURE_MODULE => RESOLVE
>>>> FIRST_MODULE_USE => POST_RESOLVE
>>>>
>>>> POST_MODULE => PRE_ACTIVATE
>>>> INSTALL => ACTIVATE
>>>> => POST_ACTIVATE
>>>>
>>>> CLEANUP => CLEANUP
>>>>
>>>> We could stop at PRE_RESOLVE. Subsequent start would go up to and
>>>> including CLEANUP. Subsequent stop would go back including
>>>> PRE_ACTIVATE.
>>>> Subsequent start as before.
>>>>
>>>> cheers
>>>> --thomas
>>>>
>>>>
>>>> --
>>>> xxxxxxxxxxxxxxxxxxxxxxxxxxxx
>>>> Thomas Diesler
>>>> JBoss OSGi Lead
>>>> JBoss, a division of Red Hat
>>>> xxxxxxxxxxxxxxxxxxxxxxxxxxxx
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> jboss-as7-dev mailing list
>>>> jboss-as7-dev(a)lists.jboss.org
>>>>
https://lists.jboss.org/mailman/listinfo/jboss-as7-dev
>>>>
>
> _______________________________________________
> jboss-as7-dev mailing list
> jboss-as7-dev(a)lists.jboss.org
>
https://lists.jboss.org/mailman/listinfo/jboss-as7-dev
--
xxxxxxxxxxxxxxxxxxxxxxxxxxxx
Thomas Diesler
JBoss OSGi Lead
JBoss, a division of Red Hat
xxxxxxxxxxxxxxxxxxxxxxxxxxxx