On Thu, Jul 15, 2010 at 8:33 PM, Emmanuel Lecharny <elecharny@gmail.com> wrote:
 Hi guys,

we have serious issues with the way we manage subentries in the server. Not that it's not working, but it's certainly not good enough for anything but a toy server.


I disagree. Adding a subentry is an administrative operation. If you do so you better be prepared to pay the price if you've done so after adding millions of entries as you point out. This is why we add subentries first then populate the server. 

The idea is that the cost is paid not during search time but during administrative changes and writes. 
 
Let me first give some heads up about what's going on.

A subentry is associated with an AdministrativePoint (AP), and defines a selection of entries which will be affected depending on the AP role. Those roles are :
- Access Control
- Collective Attributes
- SubSchema (not active atm)
- Triggers (ADS specific).

For instance, if we have a tree with a set of entries associated with a location (ie, c=France), we may define a subentry with a Collective Attribute role telling the server that every entry under the c=France branch will have a specific attribute added. We don't have then to add this attribute to *every* entry in this branch...

Anyway...

A subentry defines a selection using a filter, and a base DN for this filter to be active from.

Right now, a Subentry is attached to an AP as a (quite) normal entry, and when we add this subentry, we modidy *all* the selected entries (using the subentry filter and the base DN) will be modified to have a new attribute added. This added attribute contains a DN poiting to the associated subentry, so that when we process this entry, we can immediately know that it's associated with an AP.

So far, so good : processing an entry is fast, as we have all what we need when we have the entry. But the dark side is that if we have millions of entries, when we add an AP and a subentry, we may have to modify potentially *millions* of entries to add this attribute. Not good...

How can we improve this process ?

The idea would be to search for the APs when we process an entry, but it has to be fast. How can we do that ? Simple : we use the entry's DN and using a DN cache, we can get all the APs associated with an entry knowing its DN. It's as costly as the depth of the entry's DN. Once we have grabbed the APs, we will have to evaluate the entry to know if it's part of the selections defined by the APs' subentry. Done.

Is it costly ? Only marginaly compared to the current algorithm, as we have to lookup for the AP, when we have this list in an Attribute in the current server. But we spare the big modifications when adding - or removing/renaming/moving - a subentry.

What we just need is a APs cache and a way to process it.

This is what I will work on in the next few days if nobody objects or find a better algorithm.


I'm sorry but I have to veto this change. It's not acceptable to be taxing an LDAP server's read performance when this should be done during writes or during administrative changes.

There's a reason why I implemented it this way 5 years ago. I still think it's sound. It's a matter of when you choose to pay for the administrative change which BTW should not happen that often while reads happen all the time. I'd rather pay a lot for infrequent operations than a little bit for the most common operations.

--
Alex Karasulu
My Blog :: http://www.jroller.com/akarasulu/
Apache Directory Server :: http://directory.apache.org
Apache MINA :: http://mina.apache.org
To set up a meeting with me: http://tungle.me/AlexKarasulu