commons-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Howard M. Lewis Ship" <hls...@comcast.net>
Subject RE: [hivemind] Design comparison
Date Wed, 17 Sep 2003 17:38:42 GMT

> 
> I have been monitoring hivemind as it is similar to a 
> framework I helped design for work (not OSS). What interests 
> me is the way that different groups in Java-land are moving 
> towards the small POJO approach, and away from dreaded EJBs. 
> I just want to sketch out some features of the framework I 
> use in case it gives you some ideas.

I read a great article that, boiled down, said that "Enterprise JavaBeans" should be called
"Transactional JavaBeans" because that's all you really get.

Bloat, complexity and (due to dependencies on the app server runtime environment) limited
testablility is the problem.  All these POJOs (Picocontainer, swing, HiveMind, Avalon) are
the
reaction.

Lots of HiveMind percolated and gestated in Tapestry.


> 
> Similar to hivemind:
> --------------------
> Aims to enable a system to be written as a large number of 
> small services. Each service can call other service in 
> flexible ways. Each service is defined by an interface. The 
> interface must have only _one_method. Each implementation 
> must have no instance variables (singleton).

HiveMind services may have any number of methods.

They don't have to be singletons, though singletons are generally sufficient; the threaded
service
model allows one-instance-per-thread (but mandates that you tell HiveMind when the instances
may be
discarded).  More service models on the way.


> 
> Differences:
> -------------
> 1) The first parameter of each interface method must be a 
> Context object giving access to configuration and 
> connections. Is this IoC?

I would say not; that's configuration information that should be provided by the container.


> 
> 2) The selection of which implementation to use is performed 
> late. The selection is based on
> - data in the method arguments
> - the services in the stack calling this one
> - the configuration
> <lookupGroup interface="blah.TheInterface">
>  <lookup>
>   <caller caller="blah.SomeCallerImplementation" />
>   <process process="blah.ImplIfCallerInStack" />
>  <lookup/>
>  <lookup dataSourceType="Database">
>   <process process="blah.ImplIfParamsWantDB" />
>  <lookup/>
>  <lookup dataSourceType="File">
>   <process process="blah.ImplIfParamsWantFile" />
>  <lookup/>
> <lookupGroup/>
> I include this to give some idea of what is going on. The 
> process elements are the implementations, and which is 
> returned will depend on the current state of the system and 
> the method parameters. It acts like a big if statement.

Those almost look like AOP method introductions; this doesn't have a parallel in HiveMind
per se.

> 
> The lookup is hidden from callers by a class that simply has 
> all the interfaces as methods, performs the lookup and then 
> calls the implementation.

I'm not following this; in HiveMind services are represented by interfaces; the implementation
of
the interface may be a fabricated proxy or interceptor, or a user-supplied core implementation
(or a
fabricated core implementation).

> 
> 3) We have the concept of interceptors, again bytecode 
> generated. We use them to open and close connections. Thus 
> there is a doPre() method that opens the connection and 
> attaches it to the Context, and a doFinally(), that is called 
> as a finally block, that closes the connection. This 
> obviously simplifies the service itself, and separates system 
> logic from application logic. It is very powerful, and could 
> be worth thinking about for hivemind. (our limitation is that 
> each service can only talk to one datasource).

HiveMind allows multiple interceptors in a stack. There are examples of them performing logging
operations, and more interceptors are on the way.

> 
> 4) All our configuration is in XML resource bundles. This 
> allows locale based config for the core server behaviour 
> which we can control on a per session basis (although we 
> haven't needed to yet). We chose to separate configuration 
> from the services themselves, different to hivemind. The 
> system we use allows single Strings, Lists, Maps and raw XML 
> to be loaded from the resources by the program. There is no 
> auto-bean conversion though which is nice in hivemind 
> (although it only takes one line in our code).

Having unified module deployment descriptors enahnces the IoC aspects of HiveMind.  The framework,
as container, can 
set properties of core implementations to configurations or other services.  Proxies get in
there to
keep unecessary work from occuring
until actually needed, and to defuse cyclic dependencies.

Additionally, having interlationships between modules is very powerful; especially for pluggability
reasons.  For example, module A could define a DAO interface and service, and modules B and
C (only
one available at runtime) would contribute an implementation of the service.

HiveMind also feeds on itself ... services are used to construct other services.  BuilderFactory
is
the "IoC" engine; it constructs core implementations and can set properties of the impl to
other
services and configs.  EJBFactory creates a core impl that's a wrapper around looking up and
invoking stateless session EJBs.

--
Howard M. Lewis Ship
Creator, Tapestry: Java Web Components
http://jakarta.apache.org/tapestry
http://jakarta.apache.org/commons/sandbox/hivemind/
http://javatapestry.blogspot.com


Mime
View raw message