17 Mar
2008
17 Mar
'08
6:51 p.m.
On Mar 17, 2008, at 6:38 PM, Sergiu Dumitriu wrote:
[snip]
This allows introducing/changing ONE new functionality, by completely
replacing a whole component, but it isn't flexible enough IMHO.
No, it also allows adding actions before/after existing actions.
C. Precondition/Postcondition enforcing
The previous two use cases allow a simple ping that something has
already happened. (Actually, the listeners can also change the
passed
document, as the actual object is sent.) But I'd like to be able to
write a listener that can say "this is not allowed, please stop this
action", or "these changes are not OK, don't save this document".
So, a
precondition can be that "/save/" cannot be performed between 6PM
and
7AM, which would be a needed feature in some companies, or that "/
view/"
cannot be performed on the "SectionA" space from the IP addresses
like
192.168.2.*, which belongs to the SectionB users (these are action
events). And a postcondition can be that the XWiki.XWikiPreferences
can
only be updated by changing XWikiGlobalRights objects, other types
of
changes can only be performed by a superadmin only, or that adding
some
content that looks like spam is not allowed (these are document
events).
So, on top of the previous use cases, I'd like to add a "phase" to
the
events, which indicates if the event is sent before or after the
change
(or we can just add new events), and a new type of exception
thrown by
the listeners, which is not ignored. Right now, if a listener throws
an
exception, it is ignored. This exception, when received by the code
that
sends the notification, should stop the normal processing and
display
"action not allowed" + ex.getMessage to the user.
My first reaction is that I'm not really fan of this approach. The
way
I had planned to implement this in the V2.0 architecture is by
either:
1) replacing existing component with a different implementation. This
is when the current behavior need to be modified/replaced by another
one.
2) depending on the component, introduce new component. For example
the Action Manager in the new action component allows defining new
Actions that can be inserted before or after existing actions (or
even
replace existing actions but that's 1)).
To summarize, I would agree with your idea except that with
components
there's another way of doing it and I don't think we should have too
many ways. If you think about how you'd implement your idea you'll
end
up creating a state machine component to which actions are hooked to.
This is what the Action Manager component is meant to do, although
for
the moment I have resisted introducing a FSM since I wanted to see
the
exact need first and with 1) and 2) I thought we could already do
quite a lot (not sure what we cannot do to be honest). Using
events also allows:
- chaining several completely different "behaviors"
This is supported in the Action component architecture since you can
introduce Actions and tell where you're introducing them in the chain
of actions.
- writing one listener for all the actions, instead of
replacing
several
Actions just to introduce a check
I don't see why we would need to replace several actions to introduce
a check. You'd simply create an Action on the same mapping and make it
execute before the "main" actions for that mapping.
- defining a behavior as a Groovy script placed inside
a document, so
adding a new -- let's say -- security feature will not require
writing a
component, packaged as a jar, but just installing a wiki application,
controllable from the wiki, like the IRC bot.
This is supported too. You just need to write a Groovy class that
implements Action and register it against the component manager. The
same apply for everything. Said differently it'll be possible to
modify any behavior from groovy if we want.
I don't see a need for a FSM, as the place where
these events should
be
sent are already clear, and the old notification already sends pre-
and
post-notifications, just that all the errors are ignored.
At first I thought that the new rendering part would be enough, but
then
I realized that in order for some code to be executed, it must exist
in
the source, as a macro or something.
Then I thought that we can make several tree walkers that are
executed,
instead of one that transforms the tree into XHTML, or PDF, or RSS,
or a
different wiki markup, or whatever, but I'm not sure this is a good
idea, either.
I don't like the events approach for this (especially the fact that
the
event is used for passing data around), but for the moment it seemed a
simpler approach while waiting for the rendering engine.
I'm not sure why we need to spend time on this while waiting for the
rendering engine. As you know I'm actively working on it.
Is there something urgent? :)
In general I'm not too much in favor of tweaking the observation
mechanism to introduce the ability to replace/pre-execute/post-execute
code. Not sure why this is just a gut feeling for now. I think we need
to discuss it on a use case by use case basis. I don't see the need
for Actions and Rendering since these use case are already supported
IMO.
Thanks
-Vincent
D.
Rendering hooks
While working on the SkinExtensions and trying to write a better
#toc
macro, I was faced with the problem that there's no way a plugin can
change the rendered content at the end. Yes, it can alter it line by
line, and in theory there is an endRenderingHook sent with the
content,
but that is actually buggy, as an empty string is sent instead of
the
content, and the result is discarded. Still, even if fixed, this is
sent
only for *rendered* content, not *parsed* content. My initial idea
was
to change the PluginInterface by adding two new methods that will be
called before and after the content was generated, but that is a bad
idea, since the PluginInterface is already too big and has too many
types of methods. And since the new rendering component is not yet
available, I thought that we can also use events for this.
So, instead of writing new methods in the PluginInterface, we can
add
other type of events, which gets sent in the parsing/rendering
process.
That way, we don't need to have a plugin that implements all these
methods, and adding a new feature will require just sending a new
type
of event, instead of changing an interface (which is dangerous).
The source would be an object holding the string, which can be
modified
by the listeners. Other information could be the content document,
if
there is one, the rendering/parsing context, anything else?
Again my first reaction is that I'm no fan... :) The reason is that I
have this implemented in the new rendering component.
First any piece of code can get the parsed content of a text having
any kind of syntax (Element Blocks or DOM). Second, it's possible to
write any Macro that will manipulate this DOM and make any
modification to it. This is how the Toc Macro is going to be
rewritten
actually.
However you're right in that right now all I have planned for are
Macros and these are triggered by the macro syntax in the text to
parse/rendering. So it's a good idea to add a small FSM in the
rendering component since we already need an order in the macro
execution but also to add the ability to register macros to be
executed regardless of whether they appear in the parsed text or not. >> E. Other generic events
>> Like FlushCacheEvent, StartupEvent, ShutdownEvent, VirtualInitEvent,
>> VirtualShutdownEvent (sent when a wiki is deleted by the
>> WikiManagerEvent), Application[Added/Enabled/Disabled/Removed]Event
>> (sent by the ApplicationManager), Space[Added/Removed]Event (sent by
>> the
>> SpaceManagerPlugin), etc.
>
> +1 to those.
>
> Thanks
> -Vincent
>