tomcat-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Remy Maucherat <>
Subject Re: Proposal - Comet changes
Date Fri, 08 Sep 2006 10:13:31 GMT
Filip Hanik - Dev Lists wrote:
>> No, I don't see how filters can work. It is possible that some filters 
>> which would be wrapping the request would be ok, but most likely they 
>> would do something when the call returns (and finish what they had to 
>> do), so any attempt to use the wrapped objects would fail later on.
> The same developer that does the comet servlet sets up the mapping for 
> the filter, so I don't see a need to have to protect against this scenario.

This is not practical.

>>> 4. For interception, I think the existing valves and filters 
>>> could/should remain untouched. Interception will/should only be done 
>>> at the activation of the request
>>>   Sequential event method (such as READ) would not go through the 
>>> interceptors, just like it is today. This is essential as we can have 
>>> filters/valves that modify both incoming and outgoing content.
>> Actually, I would like to have a new type of filters (both for the 
>> container side and the application side), because things like setting 
>> up the security contexts, etc, are still likely going to be needed. 
>> This would be something simpler than valves/filters, with no mapping 
>> (since most likely, a single servlet is all that's needed to handle 
>> all the Comet traffic of an application).
>> The filter would have a filterEvent method, and the difference between 
>> the application side and the container side is that the first gets the 
>> facades, and the second the Request and Response objects (which allow 
>> access to everything). Each request should be un/wrapping as a regular 
>> filter on each invocation (of course, the wrappers can be stored in a 
>> request attribute).
> I see this as a duplicate effort, trying to recreate something that 
> already exist, here are the cons
> 1. The developers would have to learn something new, tomcat specific to 
> achieve the same thing that they could have done using filters
> 2. They would work the same way as filters, then why not use filters
> 3. They run into the same risk as you described above, ie, invalidating 
> an object at request end, can happen here as well
> 4. Assuming that there would be only *one* comet servlet per application 
> is not a safe assumption, I would never code it that way.
> 5. You lose the mapping feature, this is extremely useful, and could 
> prove to be very useful for comet servlets as well
> 6. Everyone already knows how a filter works, and they are implementing 
> a servlet, makes sense to just piggy back this functionality.

This does not make sense. They do not work the same way as regular 
filters, as every event will be intercepted, so this does not duplicate 
existing functionality. As for mapping, it can be done too if needed, 
but I think it is not that useful.

The problem for example is for a JEE server, to restore the security 
association for the thread which is processing the IO event. This can't 
be done otherwise since each event is sent using a different thread. So 
for example, the Comet servlet would not be able to access an EJB, a 
transaction, etc.

1) Yes. Different IO style, "new" API. If they can write the servlet, 
then that new filter is easier. How is that difficult ?
2) Because filters intercept the call to the "service" method, which is 
not going to occur.
3) In that case, they will not be able to code the Comet servlet either. 
Not my problem.
4) Ok. Well, it's about the same anyway. Valves (which have no mapping) 
work fine too, and I think mapping can be added later on if it turns out 
it was really needed (especially since mapping will be done only once 
per connection, it seems very practical). Personally, I would think the 
most realistic model (given it is not trivial) is to have one servlet 
take care of whatever XML protocol has been chosen, and then delegate to 
actual business logic somewhere else (= not in the servlet). That's why 
I think one servlet makes some sense.
5) You said it already. It's still something compared with a non 
existent interception model.
6) For starters, people would need to be told what they should do to 
write filters that would not break with Comet, and then find out that 
they're very limited (you can do access logging, though; oh, and 
compression, but you need to be smart ;) ).

>>> 5. StandardCometEvent could be a zero GC object if need be, if the 
>>> SecurityManager is enabled, a non reusable facade should be used
>> GC doesn't matter too much for the whole connection since it's quite 
>> long running. These objects would be discarded when the Comet 
>> connection  ends (but it would be a bit bad to start allocating too 
>> many objects for each event).
> agreed.
>>> 6. CometServlet removal - good idea.
>>> 7. The servlet should have a way of gracefully ending the session, 
>>> such as CometEvent.close() instead of just letting it timeout
>> The servlet can close the writer or OS, and the client could send the 
>> appropriate end chunk, this should work.
> that is how it is today, I think its cleaner to provide them with the 
> CometEvent.close() method and let the container handle the logistics.
> CometEvent.close may be implemented as  
> "this.getInternalResponse().getInputStream().close()" if you wish.
>>> 8. Session timeouts/invalidations. An active comet session should not 
>>> invalidate the HttpSession based on inactive time.
>>>   getAccessCount() would return >0 if there is a comet session
>>>   I am not fully kosher with this yet, still need to think about this 
>>> some more. a possibility is to have CometEvent.releaseHttpSession(), 
>>> will `--accessCount`
>> It should be ok when activity checking is enabled: endAccess will only 
>> be called when recycling the request object.
> perfect.
>>> 9. The similar problem will have to be worked out for last accessed time
>> Bleh, -1. Normally, it's still the beginning of the session access, so 
>> the beginning of the request.
> ok
>>> 10. Session replication, currently session replication is triggered 
>>> through a valve at the end of each request,
>>>    I will have to adjust clustering to support periodic replication 
>>> or add some other mechanism to make this work.
>>>    Still need to think this through.
>> So this would be done in a container side Comet event filter ;) I 
>> would not even bother with doing session replication when there's no 
>> client activity (I think anything which goes to the session when 
>> there's no client input comes from the server side, and thus can be 
>> restored as is in case of a failover). Of course, you could have 
>> periodic replication as well as an option.
> Comet event filter will not work, has the same problem I described, the 
> session can be modified async, ie a typical comet sequence is
> 1. receive data
> 2. add data to the queue
> 3. async thread reads data, processes it
> 4. async thread  writes response
> In step 3. the session could very well be modified, as the server state 
> would change.

3) is done by a worker thread under the control of Tomcat, so will go 
through filters. I don't see the point of doing 3 in other threads, 
Tomcat gives the app an independent thread to do its stuff (in case it 
needs to do more than just read the data). 4) is done by the application 
on its own. As I said, it's likely for 4) the state is not meaningful, 
or can be recreated when there's a failover (any state change would be 
simply for caching, I think).

Adding both options is going to be needed: replication triggered by a 
filter, as well as periodic replication.

I think the current API is clean enough, and it works. However, your 
proposed API fits better a real filter model, so I am willing to agree 
to doing both things, not just one, which would have little value.


To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message