[CI] Initial pass at new intro

[Samuel Madden]

On Jun 18, 2008, at 11:06 AM, Stan Zdonik wrote:

> Intro looks good.
>
> I wonder if we should play down the query rewrite part.
> At least one reviewer thought we were claiming that was novel
> and we know that BHUNT does a similar thing.  We should
> probably just say that this is how we do it to avoid having to
> change the optimizer.

Agree.

>
>
> The statement about previous work not addressing how non-hard
> dependencies are exploited is not right.  Seems like others have
> described this.
>

Perhaps this was worded wrong -- what I meant was that previous work  
doesn't really show how to determine what predicates to introduce to  
exploit such soft dependencies.  Do you disagree with that?  Clearly  
others have pointed out that it is possible to insert such predicates,  
but I don't think they discuss the data structures that determine what  
to insert -- that's our contribution.  BHUNT is the closest (with  
overflow tables), but they really don't describe how this works;  they  
mostly seem to assume constraints are truly algebraic.

> Stan
>
> Samuel Madden wrote:
>> All --
>>
>> I checked in a new intro which I think addresses the concerns about  
>> related work a little more clearly.  Please look it over and see  
>> what you think -- Alex, I believe I've captured all the ideas you  
>> guys mentioned in your mail, but I might have missed on or two.
>>
>> I'm sure we will need to iterate on this, so pls read and comment!
>>
>> There are a number of comments throughout the paper where I note  
>> things that need to fix in light of related work, etc.  Everyone  
>> should read the BHUNT and CORDS papers and understand the basic  
>> approaches since they appear to be the state of the art and are  
>> clearly pretty close to what we have done.
>>
>> I also made two other changes:
>>
>> 1) A new title, making the scheme sound less like a new index and  
>> more like a way to exploit soft function dependencies at query  
>> execution time
>>
>> 2) Changed CI to CT (correlation table), so that it doesn't seem as  
>> much like we are proposing a new index (which we aren't.)
>>
>> I think we need to clean up our story re: unclustered B+trees a  
>> little more -- they aren't really a competitive technique, just  
>> another way to exploit correlations (rather than predicate  
>> introduction) once correlations have been found.  Problem is that  
>> one could use an existing tool like CORDS and then create indices  
>> on the unclustered attributes which might be a win.
>>
>> We need to say if there is anything about c_per_u / our estimation  
>> of c_per_u that is different / better than CORDS.  If not we should  
>> cut down sections 3 and 4 and say we use the same method as CORDS  
>> so that we can dedicate more space to bucketing and multicolumn  
>> search in section 5.
>>
>> Also, there are some references to CT in the figures that I  
>> couldn't fix -- if one of you could take care of that, that'd be  
>> great.
>>
>>
>>
>> -Sam
>> _______________________________________________
>> CI mailing list
>> CI at list.cs.brown.edu
>> http://list.cs.brown.edu/mailman/listinfo/ci
>>
>
> -- 
> = 
> = 
> = 
> ======================================================================
> Stan Zdonik                            (401) 863-7648 voice
> Professor                              (401) 863-7657 fax
> Brown University                        sbz at cs.brown.edu
> Department of Computer Science P.O. Box 1910 Providence, RI 02912
> = 
> = 
> = 
> ======================================================================
> _______________________________________________
> CI mailing list
> CI at list.cs.brown.edu
> http://list.cs.brown.edu/mailman/listinfo/ci