Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Pig >> mail # dev >> CUBE/ROLLUP/GROUPING SETS syntax


Copy link to this message
-
Re: CUBE/ROLLUP/GROUPING SETS syntax
I think I'm missing something here.  The result of the "out =" line is three bags, correct?  If that's the case, the cross product you want is achieved by doing:

result = foreach out generate flatten($0), flatten($1), flatten($2)

This is not the same as CROSS, which would be expensive.

Alan.

On Jun 21, 2012, at 1:28 PM, Prasanth J wrote:

> Hello all
>
> I initially implemented ROLLUP as a separate operation with the following syntax
>
> a = ROLLUP inp BY (x,y);
>
> which does the same thing as CUBE (inserting foreach + group-by in logical plan) except that it uses RollupDimensions UDF. But the issue with this approach is that we cannot mix CUBE and ROLLUP operations together in the same syntax which is a typical case. SQL/Oracle supports using CUBE and ROLLUP together like
>
> GROUP BY CUBE(a,b,c), ROLLUP(c,d), CUBE(e,f);
>
> so I modified the pig grammar to support the similar usage. So now we can use a syntax similar to SQL
>
> out = CUBE rel BY CUBE(a,b,c), ROLLUP(c,d), CUBE(e,f);
>
> In this approach, the logical plan should introduce cartesian product between bags generated by CUBE(a,b,c), ROLLUP(c,d) and CUBE(e,f) for generating the final output. But I read from the documentation (http://pig.apache.org/docs/r0.10.0/basic.html#cross) that CROSS operator is an expensive operator and advices to use it sparingly.
>
> Is there any other way to achieve the cartesian product in a less expensive way? Also, does anyone have thoughts about this new syntax?
>
> Thanks
> -- Prasanth
>
> On May 30, 2012, at 8:10 PM, Jonathan Coveney wrote:
>
>> As far as the underlying implementation, if they all use the same
>> optimizations that you use in cube, then it can be LOCube. If they have
>> their own optimizations etc (or could), it may be worth them having their
>> own Logical operators (which might just be LOCube with flags for the time
>> being) that allows us more flexibilty. But I suppose that's between you,
>> eclipse, and your GSOC mentor.
>>
>> 2012/5/30 Prasanth J <[EMAIL PROTECTED]>
>>
>>> Thanks Alan and Jon for expressing your views.
>>>
>>> I agree with Jon's point, if the syntax contains CUBE then user expects it
>>> to perform CUBE operation. So Jon's syntax seems more meaningful and concise
>>>
>>> rel = CUBE rel BY (dims);
>>> rel = ROLLUP rel BY (dims);
>>> rel = GROUPING_SET rel BY (dims);
>>>
>>> 2 reasons why I do not prefer using SQL syntax is
>>> 1) I do not want to break into existing Group operator implementation :)
>>> 2) The syntax gets longer in case of partial hierarchical cubing/rollups
>>> For ex:
>>>
>>> rel = GROUP rel BY dim0, ROLLUP(dim1, dim2, dim3), ROLLUP(dim4,dim5,dim6),
>>> ROLLUP(dim7,dim8,dim9);
>>>
>>> whereas same thing can be expressed like
>>>
>>> rel = ROLLUP rel BY dim0,
>>> (dim1,dim2,dim3),(dim4,dim5,dim6),(dim7,dim8,dim9);
>>>
>>> Thanks Alan for pointing out the way for independently managing the
>>> operators in parser and logical/physical plan. So for all these operators
>>> (CUBE, ROLLUP, GROUPING_SET) I can just generate LOCube and use flags to
>>> differentiate between these three operations.
>>>
>>> But, yes we are proliferating operators in this case.
>>>
>>> Thanks
>>> -- Prasanth
>>>
>>> On May 30, 2012, at 4:42 PM, Alan Gates wrote:
>>>
>>>>
>>>> On May 30, 2012, at 10:43 AM, Jonathan Coveney wrote:
>>>>
>>>>> I was going to say the same thing Alan said w.r.t. operators: operators
>>> in
>>>>> the grammar can correspond to whatever logical and physical operators
>>> you
>>>>> want.
>>>>>
>>>>> As far as the principle of least astonishment compared to SQL... Pig is
>>>>> already pretty astonishing. I don't know why we would bend over
>>> backwards
>>>>> to make the syntax so similar in this case when even getting to the
>>> point
>>>>> of doing a CUBE means understanding an object model that is pretty
>>>>> different from SQL.
>>>>>
>>>>> On that note,
>>>>>
>>>>> rel = CUBE rel BY GROUPING SETS(cols);
>>>>>
>>>>> seems really confusing. I'd much rather overload the group operating