directory-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stefan Seelmann <seelm...@apache.org>
Subject Re: AdministrativePoint handling
Date Mon, 13 Dec 2010 12:28:33 GMT
Hi Emmanuel,

the lazy evaluation sounds interesting, great idea.

Just to clarify:
You are talking only about the process to determine which AP and ACIs
are applied to an entry (which is done by evaluating all AP subtree
specifications), right?
The concrete evaluation of the ACI must always be done whenever the
user accesses an entry, right?
I both is right, I wonder if the evaluation of the ACIs isn't much
more expensive then the evaluation of the subtree specifications.

Kind Regards,
Stefan



On Mon, Dec 13, 2010 at 11:41 AM, Emmanuel Lecharny <elecharny@gmail.com> wrote:
> Hi guys,
>
> I'm resuming the APs implementation. I have had a new idea this morning
> about the best way to handle them.
>
> Currently, we had devised about too possible options
>
> 1) Processing the entries while adding a new AP
> The idea is that when we add a new AP (or delete an existing one, or
> modifying an existing one), we evaluate all the underlying entries and
> update each one which is selected by the associated SubtreeSpecification by
> adding a reference to its associated AP in the entry.
> This is done once, then when a user grab an entry, we don't need to evaluate
> it again, as this work has already been done, so we accept to pay the price
> once, but it's free for every operation after this initial update.
>
> The price to pay is pretty expensive, if we consider a huge base with
> hundred of thousands entries, as we have to update many, and as an update is
> slow (we currently can process 600 per second of such updates - depending on
> the server we are running on, of course -, so for 1 000 000 entries it might
> take up to 30 minutes ).
>
> 2) Processing the evaluation on the fly
> The second idea is to evaluate the entry when a user requests it, to avoid
> paying the price for the initial update.
>
> The big problem with this approach is that if we have many APs
> SubtreeSpecifications to evaluate, this evaluation is not free, plus it has
> to occur *everytime* a user access the entry. That could slow down the
> server a lot (probably double the time it takes to return an entry).
>
> So far, we have decider to go for option #1, simply because adding an AP is
> an administrative task, which does not occur very often. We all realize that
> it has drawbacks, but it's far better than option #2 anyway.
>
> Now, I'd like to propose a third option, which might be acceptable,
> considering that we will pay a small price compared to both option #1 and
> #2.
>
> 3) Processing on the fly and store the result
> The idea is to process the evaluation when someone fetches an entry, but to
> store the evaluation result in the entry. If the fetched entry has already
> been evaluated, then no need to process it again.
>
> The direct benefit will be that we won't have this huge initial update
> required by option #1, and the entry evaluation wil be just a matter to
> check if an AT has been set instead of fully evalute the entry as in option
> #2.
>
> How will it works ? Simple : when an AP is added, a Timestamp (TS) is
> updated. This TS will be incremented every time we add a new AP. We don't
> update any entry.
> Then when a user fetch some entries, for every one of the selected entries,
> we check if it has a TS AT present. If not, then we evaluate the entry
> against the AP it depends upon, and when done, we add the last TS into the
> entry, save it and return (or not) the entry to the user.
> If the entry has a TS AT, then we have two cases
> - The TS equals the last TS : the entry has already been evaluated, and we
> can return it (or not, depending on the evaluation result)
> - The TS is below, so we reevaluate the entry.
>
> Now, how will we distinguish between an entry evaluating to true or false ?
> Easy : if the entry is selected by a SubtreeSpecification, we will have a
> ref to the associated AP in it, otherwise no.
>
> I do think that this third option has great advantages, and not only from
> the user or admin POV : the implementation will be greatly improved, as we
> don't have to deal with all the cases when handling an AP
> addition/modification/deletion.
>
> One last thing : one may object that the on the fly evaluation is not
> acceptable, as it will cost a lot for the first access to an entry. True.
> But I claim that it's the administrator to deal with this problem, and then
> its not anymore a user issue. When the admin add an new AP, he also can
> fetch all the entries below the added AP, and they will be updated on the
> fly. The users won't be impacted anymore.
>
> thoughts ?
>
> --
> Regards,
> Cordialement,
> Emmanuel L├ęcharny
> www.iktek.com
>
>

Mime
View raw message