cocoon-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ard Schrijvers" <a.schrijv...@hippo.nl>
Subject RE: Continuations consume ram, possible solutions?
Date Thu, 27 Jul 2006 13:14:58 GMT

> Hi Ard,
> switching to dev, I also think it's more appropiate.
> 
> Unfortunately I didn't have time in the last period to work a lot on
> cocoon issues, but I hope I will have some during this summer.
> 
> IMMO the main problem is not with "middle" continuations, but it's
> actually with first continuations. First continuations are the root of
> the continuation tree, and it happens often that they are created and
> never really used, for example because they are the just a 
> box aside the
> main site pages, or because they are the first step of a 
> registration or
> something else that is presented to the a user that leaves after the
> first page. Do you agree with me that also in your situation 
> great part
> of continuations are actually "first continuations"?

Yes, that is exactly the case indeed

> 
> IIRC we talked about three solutions:
> - Disk store for continuation
> - Named continuations
> - Light-continuations for entry pages
> 
> The first one is quite obvious to explain, not so simple to implement
> and does not always solve the problem. While I think is 
> rather simple to
> serialize to disk (thru EHCache or anything else) all the native
> javascript/java variables, it's a problem to do it on objects 
> : even if
> your business objects implements Serializable, that does not mean that
> they can be safely serialized and deserialized, and often serializing
> them does not necessarily mean you are freeing ram.
> 
> For example if you are using hibernate persistence, reference to your
> loaded objects are kept in the hibernate session, and 
> something similar
> happens with jpox too and so on.
> 
> Anyway, we could simply leave this problem as a programmer's concern,
> and provide a facility to store and retrieve from the disk
> continuations, only if all objects contained in it are serializable.
> Unfotunately I don't think it will be simple to have 
> serializable cocoon
> forms anyway.

I don't think cocoon gets much easier when users have to take this into account. There already
are a lot of questions about continuations/flow, and probably it will get harder when having
to take care yourself of only having serializable objects. It would probably boil down that
a few people know how to use it properly. Then again, you don't have to use it, it's an option

> 
> The second solution is very wise, and IIRC you already have something
> implemented. Using named continuations mean that every 
> continuation have
> a "name", and there can only one continuation tree with a 
> given name in
> the http session. This way, if i have a certain search page that
> everybody uses, when the user makes a search the continuation 
> is created
> and eventually its subcontinuations, but when the user goeas away from
> the page and then returns to the search page, a new continuation with
> the same name is created, which will replace the old tree, so freeing
> the ram. This way, if a typical site pattern is to do 10 searches,
> instead of having 10 continuation trees in RAM until they expire, we
> have only one.
> 
> This would not only solve "first continuations" but free the entire
> continuation trees. We could ease it a bit using the entry 
> function name
> as the continuation name, and having a "singleton" notation somewhere
> (cocoon.setSingletonContinuations()) to mark those scripts that uses
> single named continuations.

We implemented this idea recently indeed. We define wether a continuation is "unique" in its
type, we give a continuation a name, and then, when it is doubled (same name), we delete the
previous one. Though, we ran into a bug regarding this implementation recently, and I still
don't know how to solve it: 

I set the continuation to have unique=true and set its name always as follows: 
if I have a flow function in /explorer/doclisting/flow.js and the function is called "doclisting",
my continuation name would become: explorer.doclisting.flow.js.doclisting, to be sure that
it is unique. If it is tied to a specific session, I add the hashcode of the session. 

When somebody click on another folder, a new doclisting is created, and the continuation has
the same name, so the previous one can be deleted. But, when using AJAX kind of things, we
happen to use the same flow function for clicking on a document as on a folder, but does not
update all forms on the browser. But part of the forms now have continuations that are deleted!


I think this approach again can only solve some of the memory/continuations problems if you
really know what you are doing, but it is very error prone. Also solving bugs like described
above is very hard when you do not really know what is behing the idea of unique continuations.
I really doubt wether we want to go that way.

> 
> The third option aims to solve the common situation of having a flow
> that initializes some variables, sends a form or a page (creating a
> continuation) and then waits for the user to click on a 
> button, that 90%
> of time never gets pushed. This is quite common in aggregated 
> pages with
> boxes or something similar.
> 
> The idea here is NOT saving anything in the continuation, instead when
> the continuation is recalled, the function is executed again from the
> beginning, but skipping the first sendPage (since that page 
> has already
> been sent). This way we don't have any serialization issue, 
> the form is
> created, displayed and then garbage collected, and recreated when the
> user clicks the button, not displayed again since the first 
> sendPage is
> skipped, but only populated from the request.
> 
> This is maybe the simplest one to implement, but is quite 
> untidy because
> if the code in the first lines execute something heavy or 
> some business
> logic, it could not be clear why it's getting executed twice. 
> But since
> it's aimed to solve the problem of simple forms generated many times
> (like login boxes, polls, subscribe here boxes and so on), it's highly
> possible that the first lines of the flow are simply instantiating a
> bean or a document, creating a form, doing some binding and displaying
> it, and if it's really a box aside the pages of the side that code is
> already executed at every get.

This might be a lightweight sollution indeed. I suppose that you can tell the continuation
then that it is a lightweight or something? For example, lightSendPageAndWait. I think implementing
it in the original sendPageAndWait/continuation might not be backwardscompatible: the implementation
could have business logic before the continuation that actually stores something, that would
be called again in the lightweight continuation. 
> 
> Do you (plural, rest of cocoon community) have any other idea about
> this? "Continuation pollution" is actually a problem in flow
> implementations.

I do not have any ideas that you haven't pictured above :-)

Regards Ard

> 
> Simone
> 
> 
> Ard Schrijvers wrote:
> 
> >Hello Simone,
> >
> >talking about continuations, did you already find a solid 
> way to handle high traffic sites with many concurrent 
> continuations and memory useage? It bothers me a little that 
> when building sites, we have to keep track of the number of 
> continuations (we had a large site with many visitors and a 
> poll on the homepage having a continuation. This brought the 
> site down a few times).
> >
> >You mentioned serializing continuations to diskStore 
> (ofcourse, the flowscript/javaflow writer should make sure al 
> things in the cont are serializable then). Is this feasible?
> >
> >The other day I also thought about the cocoon caches having 
> this StoreJanitor trying to free memory from it when JVM is 
> low on memory. It just does not make sense to me, that this 
> is only tried regarding the cache, while currently, also 
> continuations might be the reason for a low on memory JVM. 
> >
> >Suppose, I have a healthy cache, nicely populated, and a 
> high cachehit range, but it happened to be that many many 
> continuations have been created, all long lived (5 hours), 
> and all quite large (1 continuation can be very large in 
> memory). Now, due to this continuations, JMV is low on 
> memory, implying the StoreJanitor to run, removing my cache, 
> and certainly not solving any problem.
> >
> >So I was wondering if you had some new ideas on this 
> subject...though, perhaps the dev-list is more appropriate for it.
> >
> >WDYT?
> >
> >Regards Ard  
> >
> >  
> >
> >>Hi Toby,
> >>I think you are right. What a continuation does (should do) 
> >>is dump the
> >>local variables and restore them before restarting the flow. 
> >>This means
> >>that if you write var a = 1; then create a continuation, when 
> >>you return
> >>to that continuation a should be 1 again, even if in other 
> >>continuations
> >>it has been changed to 2,3 or 4.
> >>
> >>There is surely one known limitation to this : if you say var 
> >>bean = new
> >>MyBean(); bean.setX(1); then produce a continuation, then after the
> >>continuation you call bean.setX(2), even if you go back to 
> >>the previous
> >>continuation you will find that bean.getX() == 2, because the LOCAL
> >>VARIABLE is still your bean, but it's internal state is not 
> manageable
> >>by the continuation (more formally, your local variable is a 
> >>pointer to
> >>a bean, which is correctly restored when you go back to the
> >>continuation, but the data it points to is not 
> serialized/deserialized
> >>by the continuation).
> >>
> >>But this is not your case, in this case you are setting a simple
> >>javascript variable, so it should work as you say, at least AFAIK :)
> >>
> >>Please, file a bug about it.
> >>
> >>Simone
> >>
> >>Toby wrote:
> >>
> >>    
> >>
> >>>Jason Johnston wrote:
> >>> 
> >>>
> >>>      
> >>>
> >>>>First you assign the 'useless' variable a String value, then you
> >>>>create the continuation.  When you resume the 
> continuation the first
> >>>>time, you re-assign the 'useless' variable so that it now holds an
> >>>>Array value (String.split() returns an Array).  When you 
> resume the
> >>>>continuation again, you try to call .split() on the 'useless' var,
> >>>>which is now an Array, and the error is appropriately 
> >>>>        
> >>>>
> >>thrown since an
> >>    
> >>
> >>>>Array has no such method.
> >>>>   
> >>>>
> >>>>        
> >>>>
> >>>When I resume the continuation again, I'm resuming it from before
> >>>assigning the array, so one would think that useless (which 
> >>>      
> >>>
> >>is a local
> >>    
> >>
> >>>variable) would still contain a string!
> >>>
> >>>What am I missing?
> >>>
> >>>
> >>>Toby
> >>>
> >>>-----------------------------------------------------------
> ----------
> >>>To unsubscribe, e-mail: users-unsubscribe@cocoon.apache.org
> >>>For additional commands, e-mail: users-help@cocoon.apache.org
> >>>
> >>> 
> >>>
> >>>      
> >>>
> >>-- 
> >>Simone Gianni
> >>
> >>------------------------------------------------------------
> ---------
> >>To unsubscribe, e-mail: users-unsubscribe@cocoon.apache.org
> >>For additional commands, e-mail: users-help@cocoon.apache.org
> >>
> >>
> >>    
> >>
> >
> >---------------------------------------------------------------------
> >To unsubscribe, e-mail: users-unsubscribe@cocoon.apache.org
> >For additional commands, e-mail: users-help@cocoon.apache.org
> >
> >  
> >
> -- 
> Simone Gianni
> 

Mime
View raw message