On Aug 22, 2005, at 7:29 PM, Aaron Mulder wrote:
Just pinging this thread again. Are we all agreed that we should
provide a configuration solution for M5 (as discussed in this thread
previously), that generally speaking allows each configuration to
declare
certain manageable attributes, and during GBean startup overrides the
saved values for those attributes with values it pulls out of a
user-editable XML file? I think David J identified the point to
intervene
during the loading process, so we just need any other feedback before
something is implemented.
fine with me, if it can be done quickly.
I'd also like to consider altering our J2CA WorkManager to use an
external thread pool, so that we can have one group of thread pools
that
can be managed, and then various other things (WorkManagers, EJB
network connectors, etc.) will have a reference to one of the thread
pools
defined there. Jeff was also going to look into how hard it would be
to
cause Tomcat to use a Geronimo-provided thread pool for its network
connectors instead of always using its own internal implementation.
These are good ideas, but I don't think we should hold up m5 for them.
Also, we need to resolve the GBeanName issue. I posted a
description of a conversation we had at OSCon under the thread
"GBeanNames" and no one replied. I don't fancy going ahead with the
current solution of reverting a really old patch right before every
release. I think we were close to agreement there -- perhaps we can
agree
to "store both Strings" for now, inefficient as it may be, and then we
just need to decide whether to tackle this for M5. I think it will be
a
very substantial change, as we'd be replacing ObjectName with
GBeanName in
most of the places we use an ObjectName. It's big to commit to for M5,
but perhaps even bigger to leave until a closer-to-1.0 release.
I'm in favor of keeping gbean names, making them more restrictive than
object names so every gbean name can be converted to an object name,
and having a separate query object. I don't care that much about how
many strings we store. I would prefer to get this into M5, I don't
think it will get into 1.0 if not in M5.
thanks
david jencks
Aaron
On Fri, 12 Aug 2005, David Jencks wrote:
On Aug 12, 2005, at 2:26 PM, Aaron Mulder wrote:
On Fri, 12 Aug 2005, David Jencks wrote:
What I'd like to see is:
--gbeans have regular persistent attributes and manageable
attributes
--the configuration expose the manageable attributes of the gbeans
inside
--you can only change the manageable attributes of a compiled
configuration: to change anything else you have to rebuild the
config
Just as a reminder, we already agreed on a strategy for
configurations where they will be able to be flagged as "immutable",
at
which time you could only change what we're now calling manageable
attributes. For a "mutable" configuration you could change anything.
Jeremy was going to think about what it would take to implement the
"mutable" flag on configurations, as well as adding version numbers
and so
on.
--a database abstraction/interface for the manageable attribute
values
so we can save just these values: everything else is read from the
config
--a properties file or xml implementation of the db interface so
people
can edit stuff.
I'm fine with this -- I don't think it would be very hard to
implement. I expect a GBeanInfo would have a separate Set of
manageable
properties, and the "manageable property database" would be part of
the
kernel, so when a Configuration starts a GBean it could look up any
manageable properties for the GBean and override them when loading
the
GBean. The only problem is that I'm not sure how we'd work around
properties that the GBean expects to be set in the constructor --
since
the GBean is deserialized the Configuration doesn't have the
opportunity
to call the constructor, so (for the short term) we'd probably need
to
require that manageable properties can be set via a setter before the
GBean is started (and then the Configuration could do it after the
GBean
is deserialized).
I don't understand what you mean here. GBeans aren't serialized,
GBeanData's are. If you look around line 282 of Configuration I think
there is an ideal opportunity to set the manageable attributes.
I do prefer XML over properties, and since we don't (AFAIK) want
the kernel to depend on XMLBeans, I think we could use SAX or DOM to
read
the database config. I'm OK with that since I think it would be
quite
simple -- something like:
<config>
<gbean name="...">
<property name="port" type="java.lang.Integer">8080</property>
<property name="host" type="java.lang.String">0.0.0.0</property>
</gbean>
...
</config>
Oh, and since I know David J hates maintaining XML parsing code, I
volunteer. :)
It would still be nice to keep actual xml code out of the kernel
module... I wonder if this could go in a separate module? So far the
runtime does not use any xml (except for axis)
With this, you really would have to be able to start a configuration
before you could edit values, and there are still serious problems
if
someone edits the db while the server is running, or if someone
changes
a configuration but not its version.
Well, configurations don't have versions yet, and I think that may
itself be a big change, so I think we could go ahead with the
manageable
properties but not the configuraiton versions for now.
As far as changing the DB while the server is running, we can
either say that it only takes effect on restart or we can arrange for
the
appropriate setter to be called at runtime. We can also probably
arrange
for calls to the setter at runtime to update the database.
Any other ideas on how to do this? Any ideas on how long it would
take
to implement some or all of this?
I suspect it will take longer to agree on what we want than to
implement it. :)
We just need to proclaim agreement now and start coding :-)
How far along is your configuration editing code?
thanks
david jencks
Aaron
