Re: more aggregate tests

On 14/09/10 14:54, Gregory Williams wrote:
> On Sep 14, 2010, at 4:26 AM, Andy Seaborne wrote:
>
>> Greg,
>>
>> I tried your tests : they look OK but require value-based result set testing which ARQ does not (currently) provide so it's an optical checking for now.
>>
>> The new tests weren't in the entries list and so I've committed changes to the manifest for this to CVS.
>
> Ah, missed that. Thanks.
>
>> Greg's new tests:
>>
>> Failure: Test 13 :: SUM with GROUP BY
>> Got: 5 --------------------------------
>> ------------------------------------
>> | s         | sum                  |
>> ====================================
>> | :ints     | 6                    |
>> | :mixed2   | 2.4000000000000004e0 |
>> | :mixed1   | 3.2                  |
>> | :doubles  | 32100.0e0            |
>> | :decimals | 6.7                  |
>> ------------------------------------
>> Expected: 5 -----------------------------
>> ----------------------
>> | sum    | s         |
>> ======================
>> | 2.4E0  | :mixed2   |
>> | 3.2    | :mixed1   |
>> | 3.21E4 | :doubles  |
>> | 6.7    | :decimals |
>> | 6      | :ints     |
>> ----------------------
>>
>> "E" or "e" are legal which makes comparison by lexical form tricky.
>
> Agreed. I tried to use canonical lexical form in both input and output data, but that's not guaranteed to survive a load/query.
>
>
>> =======================================
>> Failure: Test 15 :: AVG with GROUP BY
>> Got: 3 --------------------------------
>> ----------------------------------------
>> | s       | avg                        |
>> ========================================
>> | :ints   | 2.000000000000000000000000 |
>> | :mixed2 | 1.2000000000000002e0       |
>> | :mixed1 | 1.600000000000000000000000 |
>> ----------------------------------------
>> Expected: 3 -----------------------------
>> -------------------
>> | avg   | s       |
>> ===================
>> | 2     | :ints   |
>> | 1.2E0 | :mixed2 |
>> | 1.6   | :mixed1 |
>> -------------------
>>
>> More of the same. Precision issues + e/E issues.
>
> Not sure what to do about precision issues. Any thoughts? Canonical form helps, but loss of precision is a concern.

I've switched to value-based result set comparison and so I agree with 
the new tests now (not the earlier COUNT ones), except in one case:

 >> Failure: Test 13 :: SUM with GROUP BY
 >> Got: 5 --------------------------------
 >> ------------------------------------
 >> | s         | sum                  |
 >> ====================================
 >> | :ints     | 6                    |
 >> | :mixed2   | 2.4000000000000004e0 |
 >> | :mixed1   | 3.2                  |
 >> | :doubles  | 32100.0e0            |
 >> | :decimals | 6.7                  |
 >> ------------------------------------
 >> Expected: 5 -----------------------------
 >> ----------------------
 >> | sum    | s         |
 >> ======================
 >> | 2.4E0  | :mixed2   |
 >> | 3.2    | :mixed1   |
 >> | 3.21E4 | :doubles  |
 >> | 6.7    | :decimals |
 >> | 6      | :ints     |
 >> ----------------------

:mixed2 is the sum of "2.2"^^xsd:decimal and "2e-1"^^xsd:double.  So the 
decimal is converted to a double and the result added together but ARQ 
gets 2.4000000000000004e0 - there is one bit of low end noise introduced 
by the addition being done on doubles of different exponent (this is 
Java double+double doing that step).  xsd:float, not double, works.

	Andy

Received on Saturday, 18 September 2010 13:18:34 UTC