Comments at the bottom
Simon Laws wrote:
On 9/4/07, Simon Laws <[EMAIL PROTECTED] > wrote:
On 8/31/07, Simon Laws < [EMAIL PROTECTED]> wrote:
On 8/31/07, Raymond Feng < [EMAIL PROTECTED]> wrote:
In the same JVM, we now use a singleton object to keep all the
metadata for
the deployable composites running on the node. This has been proven to
be
problematic, especially in the case of Tomcat or Geronimo integration
where
applications are running in isolated address spaces. The partition is
the
portion of SCA domain visible to a given address space and its
lifecycle is
the same as an Web/JEE application (or some other modules).
Then we have the following hierarchy:
SCA domain (accross multiple nodes on the network)
--- SCA node (I assume one node per JVM and one JVM per node)
--- SCA partition (I assume one partition per address space in
the
JVM)
Thanks,
Raymond
----- Original Message -----
From: "Simon Laws" < [EMAIL PROTECTED]>
To: < tuscany-dev@ws.apache.org>
Sent: Friday, August 31, 2007 9:46 AM
Subject: Re: Rationalizing SCA Domain implementations
On 8/31/07, Raymond Feng < [EMAIL PROTECTED] > wrote:
Hi,
Comments inline.
Thanks,
Raymond
----- Original Message -----
From: "Simon Laws" < [EMAIL PROTECTED]>
To: "tuscany-dev" < tuscany-dev@ws.apache.org>
Sent: Thursday, August 30, 2007 9:13 AM
Subject: Rationalizing SCA Domain implementations
We now have a number of domain implementations in Tuscany Java
SCA
including
- The SCADomain interface
- DefaultSCADomain
- EmbeddedSCADomain
- HotUpdatableSCADomain
- DistributedDomain/Node
Covering a number of scenarios
- running a domain in a single node
- running a domain across multiple nodes
- Adding, removing, updating the contributions of a domain
through the
API
and automatically
- Activating/deactivating, starting/stopping deployable
composites
through
the API
- Starting, stopping components through the API
- Locating service in the domain through the API
etc.
There are some scenarios that we don't cover at the moment
- running multiple domains in a VM
Why does one JVM want to join multiple SCA domains? IMHO, it might
be
over-engineered.
I think this depends on the answers to you second point.
What else?
I would like to propose that we rationalize these various
implementations
down to a more manageable number. I have a simple model in my
mind of
the
building blocks we have to deal with.
1/ The node.
Is responsible for running segregated domains in a VM
Associated with zero or more domains.
I think we probably miss another layer here, the 'partition's in
the same
node. In a typical server hosting environment such as Tomcat or
Geronimo,
applications are isolated by address spaces (such as ClassLoader
for java
classes). The list of deployable composites coming from the same
address
space will form a partition. With this layer, we can better embed
Tuscany
to
the various hosting environment and provide the flexibility for
dynamic
updates.
With this in mind, the node will have an aggregate view of all the
partitions within the node.
Can you say some more about what a "partition" is in the context of
SCA?
Asking the question in a different way, what is the implication of
having
multiple "partitions" in the same domain? Is this like having
separate
nodes
but with more efficient cross node comms mechaninsm, i.e. within the
same
JVM?
2/ The domain.
Logically knows about all of the artifacts of a domain.
Associated with one or more (in the distributed case) nodes.
A local representation of the domain (the SCADomain object)
provides
the
interface to wider domain
There are some subtleties here about the timing of associating a
domain
with
a node(s) but the simple case, which we implement at the moment,
is if you start a domain, start a node, associate the two
together and
then
add contributions. The contributed components run on the node
with
which
the local domain object is associated (more complex
node/component
selection
algorithms can be imagined but we don't do this at present).
Here are some suggestions based on the interfaces from the list
of
existing
domain classes above, of the kind of things we need to be able to
do;
Domain
Create/destroy the domain based on its URI
Act of creating a domain object with a globally unique URI
means
that
it
becomes part of that domain and can scope comonent invocations in
the context of that domain.
There should be a default hot update location if we want to
maintain
that feature.
Contribution management
Add/remove contributions
Resulting composites/components sit ready to be started
Composite Management
Start/stop composite
Akin to adding a composite to the domain composite and
activating
it
Not sure how we identify a composite to be activated - by
composite
name? Currently it's done with a reference to the composite
object
Do we need to expose separate activation operations?
Component Management
Start/stop component
Add.Remove listener
Not sure how the listener from the current interface is
going to
be
used
getComponentInfo
There are some existng related component management interfaces
here
also
LocateService
public abstract <B, R extends CallableReference<B>> R cast(B
target)
public abstract <B> B getService(Class<B> businessInterface,
String
serviceName);
service name -> component name / service name
public abstract <B> ServiceReference<B>
getServiceReference(Class<B>
businessInterface, String referenceName);
We also need the domain to provide some systematic interfaces to
support
distributed operation, for example,
ServiceDiscovery
register/find service endpoint
Node
Create/destroy the node based on its URI
Add/remove an association with a domain
start/stop a nodes activity
Currently we have an approach, with the EmbeddedSCADomain, where
the
class
vends a series of management interfaces, contribution, model
building
etc.
for performing actions on the domain. This is useful as it allows
flexibility in how these management actions are implemented
without
changing
all
of the domain implementation. It would also be appropriate, in
the
future,
to expose some of these interfaces as services to allow for
remote
management of the
domain at a node.
Would welcome thoughts about this generally. Also specifically do
we
need
to
maintain the detailed interfaces provided with EmbeddedSCADomain
alongside
this.
Regards
Simon
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
Hi Raymond
I agree that the singleton is problematic.
Can you explain what you mean by "address space" in the JVM. Is it the
namespace isolation afforded to classes by their class loader?
Why would each "address space" not have a separate node? (of course this
depends on our definition of node :-)
Maybe you could say a little about how you expect the classloader
hierarchy to work in Tuscany in this situation. I note that a number of
APIs take a class loader in Tuscany but I haven't spotted of any
samples/tests that demonstrate when this faility would be used.
I will take a look at the Geronimo integration code to try and get a
better understanding of this in the morning.
Simon
Hi Raymond
I spent yesterday morning running up the Geronimo implementation. I think
I understand what you are saying now. It seems that as jars (containing
composites) are deployed to Geronimo then in the current integration
approach a new model resolver is created with a separate class loader and
the jar is contributed to the single domain that Geronimo knows about (as
defined by the plan). Will a Geronimo instance only ever know about a single
domain?
So I believe what you refer to as the partition is created by virtue of
the separate class loader that is use to resolve the resource for each
contribution.
I'm trying to work this into the interfaces I have in mind but I've been
round the houses a bit as I didn't think about this originally. I'll post
some more thoughts shortly when I can joint the dots up.
Simon
>From an implementation point of view I can enforce the restriction that one
node only deals with one domain. In the logical world the concept of a
domain spanning many runtime environments and of a physical compute engine
(JVM, process, cluster etc) supporting many domains leads to
Logical Domain n --- n Physical Compute Engine (could be JVM)
from a java implementation point of view, with this restriction, this
resolves to
Domain 1 --- 1 Node
Where Domain is the object representing a domain running on a node which is
able to find out about the other parts of the domain
and Node provides the runtime resources to this part of the domain. I still
contend that any number of Domain/Node pairs can be started within a JVM
Logical Domain 1 ---- n Domain 1 ---- 1 Node n --- n Physical Compute
Engine (could be JVM)
I think I understand what Raymond means by a partition of a node now. The
implication being that, in Java, a contribution is added to a domain with
information about not only what node it will run on but also what part of
that node. This information is currently a classpath which is used to
construct a resolver under the covers. Not sure if there is a more general
defintion of a partition key.
Domain 1 ---- 1 Node 1 --- n "Partition"
So I have changed the new domain representation as follows
A new domain implementation is based on the SCADomain interface without the
static creation methods
The node model is now internalized inside of this domain object.
It is effectively the really small runtime but with the addition of
being able to record contributions against partition keys.
The node can have a name with is used, as before, when the domain is
deployed on more than one node.
As per the first post, the Domain will vend out managers for Components,
Composites, Contributions
So if you want to create a stand alone domain, i.e a domain that relies on a
single node, as we do in many of our samples you can do.
myDomain= new DomainImpl("SomeDomainName");
myDomain.start();
myDomain.getContributionManager
("SomeContributionUrl").startContribution();
If you want to add a contribution into a partition you would do
myDomain.getContributionManager("SomeContributionUrl",
classLoader).startContribution();
In the previous case no node name was give so a default node is created for
you but it can't talk to anything else. If you want the domain to be part of
a collection of nodes you give the domain representation a node name
// Create the first domain node
domainNodeA = new DomainImpl("SomeDomainName", "nodeA");
domainNodeA.start();
domainNodeA.getContributionManager("nodeA/").startContribution();
// Create the second domain node
domainNodeB = new DomainImpl("SomeDomainName", "nodeB");
domainNodeB.start ();
domainNodeB.getContributionManager("nodeB/").startContribution();
The managers allow us to control the various aspects of the domain,
e.gcontributions, composites, components (I'm effectively using the
existing
component manager for the last bit). It gives us something to hang the hot
updateable behaviour from also but I haven't added that bit in yet.
I have code for this but it doesn't work quite yet. So any comments welcome.
Simon
[1] http://www.mail-archive.com/tuscany-dev%40ws.apache.org/msg22665.html
Some thoughts:
An SCA domain is a domain of administration. I imagine an administrator
administering an SCA domain.
A domain includes multiple nodes. A node runs software.
Let's make a node small enough to not have to be shared by two
administrators, small enough to keep it simple :)
The administrator creates/deletes nodes, installs contributions on a
node, allocates composites to a node, starts/stops/monitors a node. The
administrator gets a list of nodes in the domain and browses their
configuration.
Disclaimer: I understand that people will want to "install a
contribution on a domain and have the domain automatically find a node
or create one if necessary and download the contribution to the node and
start the node and monitor it and page me in the middle of the night if
the node is about the fail and no other node was found to take over, and
much more...". I believe that we can do all that later, once the basic
admin task building blocks I tried to list here are working.
A domain is going to be implemented as a set of Java objects, a node as
another set of Java objects. How about implementing these Java objects
without statics, singletons, threadLocals, or whatever other maps keyed
by classloaders ... then we won't have to worry about the
machine/process/JVM/partition/classloader/thread debate :)
--
Jean-Sebastien
---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
