[Zope] Archetypes leak problem? Acquisition.ImplicitAcquirerWrapper references killing eventually Zope

[Heimo Laukkanen]

First of all my apologies for long email, but this is an interesting  
problem to anyone. I appreciate any help and confirmations / views if  
anyone else can see the same.

Setup:
Python 2.3.3, Zope 2.7, Plone 2.0, Archetypes1.3b2

Problem that I am seeing now is on a service that yet does not have even a  
high load, but Zope keeps hogging steadily more and more memory, becoming  
slower and eventually coming to almost full stop - or in dire need of  
restart anyoway. Under small load this comes in few hours.

What I have seen and debugged so far is that  
Acquisition.ImplicitAcquirerWrapper are running wild in Zope. Every time  
my Archetypes-object is accessed - atleast one new reference count is  
created. Even if that object is accessed again and again.

By applying some force with ab or other tools numbers grow fast in just  
few thousand requests, below data from Zope Debug:

Acquisition.ImplicitAcquirerWrapper 127 16986 +16859
Products.PageTemplates.Expressions.SubPathExpr 79 1503 +1424
ZPublisher.HTTPRequest.HTTPRequest 31 1249 +1218
ZServer.HTTPResponse.ZServerHTTPResponse 8 1226 +1218
ZPublisher.BaseRequest.RequestContainer 5 1223 +1218

And few thousand requests later.

Acquisition.ImplicitAcquirerWrapper 127 36000 +35873
ZPublisher.HTTPRequest.HTTPRequest 31 4280 +4249
ZServer.HTTPResponse.ZServerHTTPResponse 8 4257 +4249
ZPublisher.BaseRequest.RequestContainer 5 4252 +4247
Products.PageTemplates.Expressions.SubPathExpr 79 2623 +2544
Products.PageTemplates.Expressions.PathExpr 45 1197 +1152

At this stage Zope processes have got approx. 20 percent of the full  
memory.

Now to the setup that produces following behaviour. First of all there is  
a site:

Plonesite
   SiteTree ( Archetypes object, folderish )
      intranet ( Archetypes object, folderish )

And a PythonScript called sitetreetest, which accepts obj as parameter:

---------------------------------------------

 from Products.CMFCore.utils import getToolByName

if obj is None:
     obj=context

url_tool = getToolByName(context, 'portal_url')
relative_ids = url_tool.getRelativeContentPath(obj)
o = url_tool.getPortalObject()

path_seq = []


for id in relative_ids:
     # this is much faster than restrictedTraverse and makes some sense,
     # if the user can't access the contents information, getattr() will  
fail
     # and hiding the breadcrumb because they dont have access? urk
     o = getattr(o, id)
#    o = o.restrictedTraverse(id)

     path_seq.append( o )

return path_seq

-------------------------------------------------

Script will be called from a template:

  <table border="0" cellpadding="0" cellspacing="0"           >
       <tr tal:repeat="itemi python: here.sitetreetest(here)">
        <td tal:content="python: str(itemi)">
foo
        </tD>
       </tr>
     </table>

And that is called in the context of intranet:  
/plonesite/SiteTree/intranet/template

You can actually use event SimpleFolder or other folderish archetypes  
objects, since they provide similar results.

It does not happen with Plone Folder. Code itself is similar to how Plone  
for example does breadcrumbs - so it should not itself be evil.

What also makes this interesting is that reference count for Acquisition  
wrappers starts to run wild only IF the object does not have Access  
contents information permission for anonymous user set in workflow or  
otherwise.

I've now debugged this for many, many hours and come to this point after  
excluding different other options away. Any help is appreciated as allways.

-- 
-huima