Re: [COMMITTERS] pgsql: another try at keeping AIX/ppc

I think this will cause breakage for other people.

Right now I think the order is unique to AIX for some reason.

Recent buildfarm run of gazelle should have matched the -0 variant
(cube_1.out), but did not.

Also, what is worrisome is that there is an ORDER BY on the result set
that is failing on AIX, so I think there might be a deeper problem
lurking here.

Thanks,
-rocco

> -----Original Message-----
> From: pgsql-committers-owner(at)postgresql(dot)org
> [mailto:pgsql-committers-owner(at)postgresql(dot)org] On Behalf Of
> Andrew Dunstan
> Sent: Thursday, July 27, 2006 2:39 PM
> To: pgsql-committers(at)postgresql(dot)org
> Subject: [COMMITTERS] pgsql: another try at keeping AIX/ppc
> happy on cube test.
>
>
> Log Message:
> -----------
> another try at keeping AIX/ppc happy on cube test.
>
> Modified Files:
> --------------
> pgsql/contrib/cube/expected:
> cube_1.out (r1.4 -> r1.5)
>
> (http://developer.postgresql.org/cvsweb.cgi/pgsql/contrib/cube
/expected/cube_1.out.diff?r1=1.4&r2=1.5)

---------------------------(end of broadcast)---------------------------
TIP 5: don't forget to increase your free space map settings

Indeed. I am going to revert it.

The trouble is we currently have several orthogonal variations, which is
a worry on its own:

. negative zero
. exponent format, and
. result ordering

Looking closer, the result ordering certainly does seem odd, as you say.
Surely with ORDER BY it should be deterministic. What is the rule that
describes the expected ordering of cube objects?

cheers

andrew

Rocco Altier wrote:

>I think this will cause breakage for other people.
>
>Right now I think the order is unique to AIX for some reason.
>
>Recent buildfarm run of gazelle should have matched the -0 variant
>(cube_1.out), but did not.
>
>Also, what is worrisome is that there is an ORDER BY on the result set
>that is failing on AIX, so I think there might be a deeper problem
>lurking here.
>
>Thanks,
> -rocco
>
>
>
>>-----Original Message-----
>>From: pgsql-committers-owner(at)postgresql(dot)org
>>[mailto:pgsql-committers-owner(at)postgresql(dot)org] On Behalf Of
>>Andrew Dunstan
>>Sent: Thursday, July 27, 2006 2:39 PM
>>To: pgsql-committers(at)postgresql(dot)org
>>Subject: [COMMITTERS] pgsql: another try at keeping AIX/ppc
>>happy on cube test.
>>
>>
>>Log Message:
>>-----------
>>another try at keeping AIX/ppc happy on cube test.
>>
>>Modified Files:
>>--------------
>> pgsql/contrib/cube/expected:
>> cube_1.out (r1.4 -> r1.5)
>>
>>(http://developer.postgresql.org/cvsweb.cgi/pgsql/contrib/cube
>>
>>
>/expected/cube_1.out.diff?r1=1.4&r2=1.5)
>
>---------------------------(end of broadcast)---------------------------
>TIP 5: don't forget to increase your free space map settings
>
>---------------------------(end of broadcast)---------------------------
>TIP 1: if posting/reading through Usenet, please send an appropriate
> subscribe-nomail command to majordomo(at)postgresql(dot)org so that your
> message can get through to the mailing list cleanly
>
>
>

I just updated to the latest HEAD, so I assume I have the cube_1.out
that you changed. If you look at the order of the results expected in
cube.out and cube_1.out, they are different. So I don't think we have a
problem with the code, we just need to fix the ordering in cube_1.out.

The same problem appears in cube_2.out -- is this used by any platform?

Josh

Andrew Dunstan wrote:
>
> Indeed. I am going to revert it.
>
> The trouble is we currently have several orthogonal variations, which is
> a worry on its own:
>
> . negative zero
> . exponent format, and
> . result ordering
>
> Looking closer, the result ordering certainly does seem odd, as you say.
> Surely with ORDER BY it should be deterministic. What is the rule that
> describes the expected ordering of cube objects?
>
> cheers
>
> andrew
>
>
> Rocco Altier wrote:
>
>> I think this will cause breakage for other people.
>>
>> Right now I think the order is unique to AIX for some reason.
>>
>> Recent buildfarm run of gazelle should have matched the -0 variant
>> (cube_1.out), but did not.
>>
>> Also, what is worrisome is that there is an ORDER BY on the result set
>> that is failing on AIX, so I think there might be a deeper problem
>> lurking here.
>>
>> Thanks,
>> -rocco
>>
>>
>>
>>> -----Original Message-----
>>> From: pgsql-committers-owner(at)postgresql(dot)org
>>> [mailto:pgsql-committers-owner(at)postgresql(dot)org] On Behalf Of Andrew
>>> Dunstan
>>> Sent: Thursday, July 27, 2006 2:39 PM
>>> To: pgsql-committers(at)postgresql(dot)org
>>> Subject: [COMMITTERS] pgsql: another try at keeping AIX/ppc happy on
>>> cube test.
>>>
>>>
>>> Log Message:
>>> -----------
>>> another try at keeping AIX/ppc happy on cube test.
>>>
>>> Modified Files:
>>> --------------
>>> pgsql/contrib/cube/expected:
>>> cube_1.out (r1.4 -> r1.5)
>>> (http://developer.postgresql.org/cvsweb.cgi/pgsql/contrib/cube
>>>
>> /expected/cube_1.out.diff?r1=1.4&r2=1.5)
>>
>> ---------------------------(end of broadcast)---------------------------
>> TIP 5: don't forget to increase your free space map settings
>>
>> ---------------------------(end of broadcast)---------------------------
>> TIP 1: if posting/reading through Usenet, please send an appropriate
>> subscribe-nomail command to majordomo(at)postgresql(dot)org so that your
>> message can get through to the mailing list cleanly
>>
>>
>>
>
>
> ---------------------------(end of broadcast)---------------------------
> TIP 3: Have you checked our extensive FAQ?
>
> http://www.postgresql.org/docs/faq

--
Joshua Reich
Finance and Corporate Development
ROOT Exchange, A Division of ROOT Markets
601 W. 26th St. / Suite 1500
New York, NY 10001
W - (212) 645 6320 x 7101
M / T - (646) 427 7959
E - josh(at)rootexchange(dot)com

Joshua Reich wrote:

> I just updated to the latest HEAD, so I assume I have the cube_1.out
> that you changed. If you look at the order of the results expected in
> cube.out and cube_1.out, they are different. So I don't think we have
> a problem with the code, we just need to fix the ordering in cube_1.out.
>

The version where I reverted the change is 1.6.

There is apparently a code problem somehow, since we know that AIX/PPC
at least orders things differently on those 2 tests.

I have looked at the new code and can't see any obvious reason for
ordering differences.

Maybe we need to try to stress test the comparison routines a bit to
make sure they really are deterministic.

> The same problem appears in cube_2.out -- is this used by any platform?
>
>

We will deal with that one in due course.

cheers

andrew

Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
> Maybe we need to try to stress test the comparison routines a bit to
> make sure they really are deterministic.

Eyeball inspection shows that cube_cmp is wrong: it's doing
PG_RETURN_INT16 where it should say PG_RETURN_INT32.

As best I can tell, the committed expected file is actually wrong.
I'm going to patch both cube.out and cube_1.out to match what I
get after fixing the function, and we'll see where that takes us.

I notice that cube_2.out hasn't been updated. Was that intentional?

regards, tom lane

Tom Lane wrote:
> Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
>
>> Maybe we need to try to stress test the comparison routines a bit to
>> make sure they really are deterministic.
>>
>
> Eyeball inspection shows that cube_cmp is wrong: it's doing
> PG_RETURN_INT16 where it should say PG_RETURN_INT32.
>

ouch, good catch!

> As best I can tell, the committed expected file is actually wrong.
> I'm going to patch both cube.out and cube_1.out to match what I
> get after fixing the function, and we'll see where that takes us.
>
> I notice that cube_2.out hasn't been updated. Was that intentional?
>

Well, I was going to wait to see a buildfarm member that needed the
alternative exponent representation, and use the regression diff as a
template

cheers

andrew

Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
> Tom Lane wrote:
>> I notice that cube_2.out hasn't been updated. Was that intentional?

> Well, I was going to wait to see a buildfarm member that needed the
> alternative exponent representation, and use the regression diff as a
> template

I've had pretty good success with updating expected files for platforms
I can't test by using patch to apply the same update to them as I have
for the version I can test. That works unless the diff itself contains
platform dependencies. We'll soon find out whether that's true for this
particular case ...

regards, tom lane