Hi Rajith,
I have already implemented the session affinity based loadbalancing. It
can bind a session id (e.g.soap session, RM session) to an endpoint, so
that all successive messages from that session will be directed to the bound
endpoint. In our new endpoint model, this endpoint can also
be a failover endpoint. Therefore, a session can be bound to a failover
group, which contains one master node (at a given time) and set of
backup nodes. So if we consider the failover group as an Axis2 cluster, we
can blindly replicate the state of the master node to backup nodes
without considering the concurrency issues.
I have to change my code a bit to suite the new endpoint model and NIO. I
will commit the new code ASAP.
Chathura
On 3/10/07, Rajith Attapattu <[EMAIL PROTECTED]> wrote:
Chathura, Paul,
Can these load balances identify affinity based on client properties (like
service group id or some other identifier) to make sure requests from a
particular client (or clients accessing the same service group) is always
redirected to a particular node. This will help a lot in the axis2 case, so
we can do cheap clustering w/o worrying about concurrent access to the
replicated contexts. In this case we can replicate to backup nodes blindly
bcos the load balancer will always ensure one primary.
I guess this could be another policy :)
Sorry if I am causing u trouble .
Rajith
On 2/21/07, Chathura Ekanayake < [EMAIL PROTECTED]> wrote:
>
> Hi All,
>
> Another advantage I see in the new endpoint model is that we can nest
> load balance and failover groups inside each other. One good use case for
> this is that we can load balance over set of failover groups to acheive
> scalability as well as high availability.
>
> So I propose following changes to endpoints and to the send mediator to
> support this model.
>
> We can have a Endpoint interface which is implemented by all the endpoints.
>
> public interface Endpoint {
>
> public void send(MessageContext synapseMessageContext);
> }
>
> send method sends the message according to the given endpoint object. For
example a basic endpoint would send the message to the actual endpoint address and
a load balance endpoint would select another endpoint according to a load balance
policy and call its send method. So we can achieve the nested structure.
>
>
> Send mediator has a single endpoint object in which there can be a hierarchy
of endpoints. When mediating, the send mediator just calls
> the send method of its endpoint object.
>
> Each endpoint has its own endpoint factory which implements the
EndpointFactory interface.
>
>
> public interface EndpointFactory {
> public Endpoint createEndpoint(OMElement endpointConfig);
> }
>
> SendMediatorFactory will call above factory method for its immediate child
endpoint. And each EndpointFactory will call factory methods for their child
endpoints except for leaf level endpoints (address and WSDL endpoints). This is
similar to the AbstractListMediatorFactory.
>
>
> I think we can achieve the nested endpoint model cleanly using this design.
If this endpoint model and the desing is ok, I will start working on it.
>
> Comments...
>
>
> Chathura
>
>
>
> On 2/22/07, Paul Fremantle <[EMAIL PROTECTED] > wrote:
> >
> > Chathura has written some load-balancing code for Synapse, but it
> > seems like the current model for endpoints needs a little refactoring
> > to sit well with it.
> >
> > Currently an endpoint is "just" one endpoint. In the config language
> > (http://ws.apache.org/synapse/Synapse_Configuration_Language.html) it
> > says you can have several endpoints in a <send>..</send> but we've
> > never implemented that.
> >
> > If we did it that way, then you couldn't use the load-balancing with a
> > proxy without defining a sequence, because the proxy just takes a
> > single endpoint reference.
> >
> > So it seems it makes more sense to have the <endpoint> tag refer to a
> > "logical" endpoint. A logical endpoint could be defined in lots of
> > ways. For example, it could actually be a load-balancing group, or a
> > failover pair.
> >
> > So here is a potential syntax:
> >
> > <endpoint name="string">
> > (
> > <address uri="http://"; [format="soap|soap11|soap12|pox"]>
> > <enableRM policy=" "/>?
> > <enableSec [policy="key"]/>?
> > <enableAddressing/>?
> > .. extensibility ..
> > </address> |
> >
> > <wsdl [uri="http://..";] [service="qname"] [port/endpoint="qname"]>
> > <wsdl:definition>...</wsdl:definition>?
> > <wsdl20:description>...</wsdl20:description>?
> > <enableRM policy=" "/>?
> > <enableSec [policy="key"]/>?
> > <enableAddressing/>?
> > .. extensibility ..
> > </wsdl> |
> >
> > <loadbalance policy="round-robin">
> > <endpoint ../>+
> > </loadbalance> |
> >
> > <failover policy="exponential-retry">
> > <endpoint ../>+
> > </failover> |
> >
> > .. extensibility .. )
> > </endpoint>
> >
> > So in this model, we can define an endpoint as either an address, or
> > using a WSDL, or as a load-balance group, or failover group. There
> > could be other types (maybe using MEX?).
> >
> > Chathura has some cool implementation thoughts to go with this, which
> > I hope he will send as well (hint, hint).
> >
> > Paul
> >
> > --
> > Paul Fremantle
> > VP/Technology, WSO2 and OASIS WS-RX TC Co-chair
> >
> > http://bloglines.com/blog/paulfremantle
> > [EMAIL PROTECTED]
> >
> > "Oxygenating the Web Service Platform", www.wso2.com
> >
> > ---------------------------------------------------------------------
> > To unsubscribe, e-mail: [EMAIL PROTECTED]
> > For additional commands, e-mail: [EMAIL PROTECTED]
> >
> >
>
