jackrabbit-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gregory Block <gbl...@ctoforaday.com>
Subject Re: problem with index not being updated
Date Fri, 13 Jan 2006 16:56:40 GMT

On 13 Jan 2006, at 13:10, Cyrus Daboo wrote:
> The text filter is very specific to the webapp (it does some custom  
> indexing of the iCalendar data - not a straight text extraction as  
> other text filters do) and it has a dependency on iCalendar  
> libraries for parsing. Thus a whole bunch of stuff, not just the  
> text filter, would have to be moved up to the tomcat class loader -  
> this seems like a bad layering violation to me.

It is - but the problem is more fundamental.

Fundamentally, you've got two paths you'll end up on:

1) You keep the core outside the war classloader, and the filter in  
the system classloader.  Because the core application is using shared  
classes, references to classes that aren't available in the WAR  
classlaoder is available to other WARs - however, they are invalid  
when used there.  Keep it in the WAR classloader, and you risk  
creating a dirty environment for other WARs.

2) You move the core into the WAR with the rest of it - but now you  
need to ensure that your shutdown hooks are getting called by the war  
to shut down the environment properly.  You can either use HiveMind  
or some other dependency injection system to perform the shutdown on  
registry shutdown (and call shutdown from a filter, for example) or  
use one of the servlet unloads to do so.  But now you can't share  
anything between instances, and nothing's visible to other  
applications, some of which may be SSO.

There's no good way around this - we have the same problem in the  
Castor project; when the core is loaded from common, classes loaded  
into it become a part of the system that's visible, if someone tries  
hard enough, to other classloaders in other wars, and do cause  
layering violations.

There's a design-around, I guess, though:  Ensuring that all access  
to the core is done through a 'context', and the context retains all  
of the classloaded filters, etc.; releasing that context can then  
release the classloader-dependent, environment-specific stuff, while  
preventing the layering problems.  On the other hand this, by design,  
prevents transparent sharing of the global context, as a global  
context no longer exists - applications are responsible for knowing  
when a context is shareable based on the stuff that's loaded into it  
and their own use of the context.

If you won't move to 2), you have to make 1) tenable - that means  
that when your application shuts down, you need to find a way to  
basically either unregister everything you registered which was war- 
specific, or return the system to (either an initialized or a  
shutdown) state where none of those classes are loaded.

Not uncommon, no - but not easily avoidable when people violate the  
layering by passing class implementations in the war out to something  
that runs, and continues to run beyond the war's unload, in a set of  
other threads.  Ultimately, the WAR was responsible for creating the  
layering violation - it's the WAR's job of doing whatever's necessary  
to make unload possible.

That may mean API for de-registering the stuff you've added, or it  
may mean forcing a shutdown of the jackrabbit container.  Them's the  
breaks.


Mime
View raw message