Alec Mitchell wrote:
Howdy CMFers,

So, Sidnei has been plugging away at the "AT reindexes things an obscene number of times" issue today, and appears to have fixed many of the AT triggered indexing redundancies. There are however still a few places in CMF where some cataloging redundancy might be avoided. One obvious place is during object creation, where the following happens:

*) TypesTool.constructInstance() is triggered
**) A _setObject call results in CMFCatalogAware.manage_afterAdd() which triggers a full indexObject().
    *) This is shortly followed by TypesTool._finishConstruction()
        *) Which calls CMFCatalogAware.notifyWorkflowCreated()
            *) Which in turn calls WorkFlowTool._reindexWorkflowVariables()
**) Which does a CMFCatalogAware.reindexObject([idxs]) on workflow specific variables (with a full metadata update) *) And calls CMFCatalogAware.reindexObjectSecurity() which reindexes the object only on the security index, and doesn't touch metadata. **) TypesTool._finishConstruction() then does another CMFCatalogAware.reindexObject().

So we have two full reindexes, and three metadata updates. The last reindex appears to be there only to catch the change to 'portal_type' in _finishConstruction. So, this final reindexObject, might safely be changed to reindexObject(['portal_type', 'Type']),

This was the case in my initial code, but Yuppie changed it:
http://svn.zope.org/trunk/CMFCore/TypesTool.py?rev=35903&r1=35864&r2=35903
I don't remember what the reason was, though I believe it was discussed a bit at the time on the lists.

though the possibility exists that other indexed attributes added by 3rd parties may depend on the value of portal_type (say, I use an autogenerated Title which includes the Type). Additionally, almost immediately before this last reindexObject call, another reindexObject call has happened in notifyWorkflowCreated, which included a full catalog metadata update. As a result, updating the catalog metadata here is certainly redundant. Unfortunately, the CMFCatalogAware.reindexObject method provides no means of avoiding the duplicate metadata update, though it would be trivial to add and to use here.

But as you realize, there is a problem when you have metadata computed using methods. As exemplified by portal_type / Type, just because one attribute is modified doesn't mean only one metadata (or index for that matter) is changed.

Another option suggested by Sidnei on IRC, which would avoid the potential issues with limiting the variables indexed in the final reindex. Would be to let CMFCatalogAware.manage_afterAdd know (presumably via some state variable) that it is being invoked through constructInstance/invokeFactory, in which case it could safely skip the initial indexing and allow _finishConstruction to take care of indexing the object fully on it's own at the end.

That's certainly a good hack. There are several ways to do it, either with a thread-local variable, or in the request, or by walking the stack's locals to check for a __dont_index__ attribute... You'd have to bench, but a thread-local variable is probably the fastest. You want to store a set of objects whose indexing should be skipped.

In the long term we will probably be better served by delaying all indexing to transaction boundaries, though it will be a fair bit harder to implement, and may irk some developers who depend on immediate changes to the catalog on reindex.

As Julien said it's not very hard to implement, it's just that there are application changes to consider. Still, there's agreement that CMF should move in that direction, I can provide patches taken from the CPS implementation. (And it requires Zope 2.8/ZODB 3.4 of course.) Some of the framework should pushed into Zope itself.

Florent

--
Florent Guillaume, Nuxeo (Paris, France)   Director of R&D
+33 1 40 33 71 59   http://nuxeo.com   [EMAIL PROTECTED]
_______________________________________________
Zope-CMF maillist  -  Zope-CMF@lists.zope.org
http://mail.zope.org/mailman/listinfo/zope-cmf

See http://collector.zope.org/CMF for bug reports and feature requests

Reply via email to