Andrie,
Let's separate the two issues. First, I think you're ok with the
minor change passing the properties map down to the register
transformer. I absolutely need this and I don't believe you have an
issue with this, correct?
Second, the concurrent Maps issue. Let me remove this from the
patch for now so Gemini can move ahead and let's you and I look at
the larger issue of initialization and concurrency in the
initialization code. I've worked around the current EclipseLink
expectation of a single initialization pass but am not happy with
it. It'll work for now but EclipseLink initialization should be
more flexible, especially how it obtains classloaders.
Once I have the Gemini code committed we can review what I needed to
do to workaround the current limitations and perhaps use this
example to drive some changes into EclipseLink?
Shaun
On 12/7/2010 5:02 PM, Andrei Ilitchev wrote:
On 12/07/2010 4:29 PM, Shaun Smith wrote:
Hi Andrei,
Unfortunately you haven't seen the Gemini JPA code that
relies on
these EclipseLink changes because Mike Keith is waiting for
these
changes to show up in EclipseLink before we apply the Gemini
updates.
Please post these changes - otherwise it's impossible to review
the
current patch.
SE Initialization is a pretty sensitive (and specialized) code,
let's
be careful..
On 12/7/2010 4:15 PM, Andrei Ilitchev wrote:
JPAInitializer: why to add an extra parameter to
registerTransformer
(Map properties)? The only use of the existing parameter
puInfo is to
print out puName - for FINER logging.
In Gemini OSGi JPA I need to pass properties down to the
transformer,
specifically the Bundle name and version.
EntityManagerFactoryProvider: both initialEmSetupImpls and
initialPuInfos are used only in JavaSE case (not OSGi) - why
both their
class and initial values are changed? They should not be.
They are referenced in PersistenceProvider which is not just
SE. The
Gemini code path will hit them.
Again, let's look at Gemini code path...
HashMap is used so that get doesn't lock the map - Hashtable's
locking
get method would cause performance deterioration.
So we use a highly concurrent map, but truth be told this code
is not
highly contended.
Please point exactly at the code where the problem seems to be.
There
is no concurrency concerns for initialPuInfos - it's used (if
used at
all) only once during initialization;
Again, referenced in PersistenceProvider and in Gemini
initialization
happens as bundles come and go.
Possible concurrency issues for initialEmSetupImpls resolved
by
synchronized (EntityManagerFactoryProvider.emSetupImpls)
statement in
PersistenceProvider (one for put, another for remove).
I think ConcurrentHashMap is a more concurrent solution.
What is "more concurrent solution"? Is it more efficient? Or do
you see
a concrete problem with the existing code - if so please post this
code.
Initial values should stay null for both maps, too:
neither of the maps is used outside of JavaSE case - no reason
to
always initialize them.
Again, PersistenceProvider and the Gemini code is not JavaSE,
it's
multi-threaded OSGi. It's easier to initialize them statically
which
is thread safe than to write multi-thread aware initialization
code.
Even
in JavaSE case initialPuInfos only used if the javaagent is
used -
initializing it application managed case on app. server would
mean that
changes made to redeployed persistence unit would be ignored.
Gemini provides an OSGi "Agent" that may run multiple times for
different bundles to perform initialization and support
weaving.
But could you clarify the issue with redeployment?
initialPuInfos was introduced as a performance enhancing trick for
a
single very specific case - JavaSE using javaagent.
In JavaSE case that doesn't use javaagent there is a potential for
redeployment: say, it's an application on app. server, it could be
undeployed, then redeployed with possibly altered classes,
persistence.xml etc.
If the original puInfo is cached in initialPuInfos, then the new
puInfo
will be ignored.
If bundles come and go, I guess OSGi would need
initialPuInfos==null.
Shaun
On 12/07/2010 3:15 PM, Shaun Smith wrote:
Bug 312947 - Allow
EclipseLink JPA to work with Gemini in OSGi
https://bugs.eclipse.org/bugs/show_bug.cgi?id=312947
Minor adjustments to JPA initialization to work with Gemini
JPA. This
revised patch does not remove what seemed to be dead
temploader code as
James advised is still necessary to support weaving legacy
apps using
sessions.xml.
--
Shaun
Smith | Principal Product Manager
Phone: +19055023094
Oracle Server Technologies,
Oracle
TopLink
ORACLE Canada | 110 Matheson Boulevard West, Suite 100,
Mississauga,
Ontario | L5R 3P4
 Oracle is
committed to developing practices and products that help
protect the
environment
_______________________________________________
eclipselink-dev mailing list
eclipselink-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/eclipselink-dev
_______________________________________________
eclipselink-dev mailing list
eclipselink-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/eclipselink-dev
--
Shaun
Smith | Principal Product Manager
Phone: +19055023094
Oracle Server Technologies,
Oracle
TopLink
ORACLE Canada | 110 Matheson Boulevard West, Suite 100,
Mississauga,
Ontario | L5R 3P4
 Oracle
is committed to developing practices and products that help
protect the
environment
_______________________________________________
eclipselink-dev mailing list
eclipselink-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/eclipselink-dev
_______________________________________________
eclipselink-dev mailing list
eclipselink-dev@xxxxxxxxxxx
https://dev.eclipse.org/mailman/listinfo/eclipselink-dev
--
Shaun Smith | Principal Product Manager
Phone: +19055023094
Oracle Server Technologies, Oracle
TopLink
ORACLE Canada | 110 Matheson Boulevard West, Suite 100,
Mississauga, Ontario | L5R 3P4
Oracle is committed to developing practices and
products that help protect the environment
|
