[jboss-as7-dev] AS7-4007 missing Infinispan dependency for clustered JPA second level cache

[David M. Lloyd]

On 03/06/2012 01:02 PM, Galder Zamarreño wrote:
> On Mar 6, 2012, at 6:31 PM, Paul Ferraro wrote:
>> ----- Original Message -----
>>> From: "Jason T. Greene"<jason.greene at redhat.com>
>>> To: "David M. Lloyd"<david.lloyd at redhat.com>
>>> Cc: "Galder Zamarreño"<galder at redhat.com>, "Paul Ferraro"<pferraro at redhat.com>, "Bela Ban"<bban at redhat.com>,
>>> "infinispan -Dev List"<infinispan-dev at lists.jboss.org>, hibernate-dev at lists.jboss.org,
>>> "jboss-as7-dev at lists.jboss.org Development"<jboss-as7-dev at lists.jboss.org>
>>> Sent: Tuesday, March 6, 2012 10:30:11 AM
>>> Subject: Re: [jboss-as7-dev] AS7-4007 missing Infinispan dependency for clustered JPA second level cache
>>>
>>> On 3/6/12 9:28 AM, David M. Lloyd wrote:
>>>> On 03/06/2012 09:21 AM, Galder Zamarreño wrote:
>>>>> This reminds me that we had a discussion about this a few months
>>>>> back: http://goo.gl/DJLhB
>>>>>
>>>>> At the time, it wasn't clear whether you can use
>>>>> ModularClassResolver in a non-module env. Seems like it's not
>>>>> possible.
>>>>>
>>>>> So, one option is to have a class resolver lookup class, or
>>>>> similar so that AS7 can pass us the right ClassResolver.
>>>>>
>>>>> There might a quicker option here though:
>>>>> - Extend VersionAwareMarshaller (http://goo.gl/bu8CF) and provide
>>>>> a variant of JBossMarshaller (http://goo.gl/o9Ro3) in Infinispan
>>>>> (this requires JBossMarshaller to be made non-final)
>>>>> - Override JBossMarshaller.inject(), call super.inject() and then
>>>>> call baseCfg.setClassResolver(…)
>>>>> - This would require that the VersionAwareMarshaller extension be
>>>>> passed the ModuleLoader that ModularClassResolver requires.
>>>>> - Configure Infinispan with this new marshaller, via
>>>>> SerializationConfiguration. Thankfully you can now pass an
>>>>> instance of the marshaller to it, so no need to do any magic
>>>>> ModuleLoader lookup from the VersionAwareMarshaller extension.
>>>>>
>>>>> Scott, other than the change in JBossMarshaller to be made
>>>>> non-final, the rest can built on AS7. Wanna have a go? Ping me on
>>>>> IRC if you need help.
>>>>>
>>>>> Cheers,
>>>>>
>>>>> p.s. Silly question, how come this issue is not present in HTTP
>>>>> session replication use case or EJB3 SFSB?
>>>>
>>>> It actually is present anywhere we're using Infinispan (I think).
>>>> It
>>>> just hasn't blown up yet anywhere public (in this case it's a 90%
>>>> effective solution versus a 100% effective solution).
>>>
>>> Session replication does setup a custom marshaller, so it does in
>>> fact
>>> work. I think the issue is just that the hibernate-infinispan, which
>>> is
>>> part of the hibernate project set, and not AS code does not do this,
>>> and
>>> it doesn't expose a way to do it.
>>
>> Sort of...
>>
>> To summarize:
>> The fundamental issue is that Infinispan uses the same marshaller instance for all caches within a cache manager - which doesn't map cleanly to our use case, which uses a cache instance per application - thus requires cache-grained marshalling configuration.
>
> Well, this is not exactly correct since 5.0. Internally, we differentiate between a global marshaller and a per-cache marshaller. What's the difference? The classloader with which they're configured.
>
> Since 5.0, you can configure the classloader to use per cache. You can even configure the classloader per invocation with cache.with(cl)….
>
> If that's not helping you, we were wasting our time when we did this...

Well.... :)  (see below)

>
>> To work around this, we typically store MarshalledValues in the cache - which are marshalled/unmarshalled using a marshalling configuration specific to the application (e.g. via a ModularClassResolver using the ModuleLoader of the deployment unit).
>> https://github.com/jbossas/jboss-as/blob/master/clustering/api/src/main/java/org/jboss/as/clustering/MarshalledValue.java
>> https://github.com/jbossas/jboss-as/blob/master/clustering/api/src/main/java/org/jboss/as/clustering/SimpleMarshalledValue.java
>> https://github.com/jbossas/jboss-as/blob/master/clustering/api/src/main/java/org/jboss/as/clustering/HashableMarshalledValue.java
>
> Isn't a class resolver and a class loader, functionality wise, doing the same thing? I wonder if a custom classloader could not be built that delegates to a ModularClassResolver...

No, not really.  A class loader loads a class, given a name.  But a 
class resolver loads a class given a name *and* stream information.  The 
modular class resolver reads the stream information to know which class 
loader houses the class in question.  This is critically important 
because it's possible (common even) for more than one class to exist in 
an AS instance with the same name.  And there is no single class loader 
which has visibility to all classes which could potentially be stored in 
a cache.

Yes, accepting a class loader to use is a powerful feature.  However it 
*should* just be a convenience abstraction over a more fundamental 
feature which allows a class resolver to be set.

>> So, essentially Infinispan itself only ever has to marshal/unmarshal a byte[] wrapper - so the AS has full control over the marshalling process.
>>
>> I would recommend that the 2LC do something similar, and include a mechanism for providing a MarshallingConfiguration per persistence unit.
>
> Possibly…
>
> --
> Galder Zamarreño
> Sr. Software Engineer
> Infinispan, JBoss Cache
>


-- 
- DML