Re: pika failing since the per-column collation patch

Hi,

Since the per-column collation patch went in, pika (NetBSD 5.1/mips) started failing consistently with this diff:

*** /home/pgbuildfarm/workdir/HEAD/pgsql.15101/src/test/regress/expected/polymorphism.out Sat Feb 12 02:16:07 2011
--- /home/pgbuildfarm/workdir/HEAD/pgsql.15101/src/test/regress/results/polymorphism.out Sat Feb 12 09:10:21 2011
***************
*** 624,630 ****

-- such functions must protect themselves if varying element type isn't OK
select max(histogram_bounds) from pg_stats;
! ERROR: cannot compare arrays of different element types
-- test variadic polymorphic functions
create function myleast(variadic anyarray) returns anyelement as $$
select min($1[i]) from generate_subscripts($1,1) g(i)
--- 624,630 ----

-- such functions must protect themselves if varying element type isn't OK
select max(histogram_bounds) from pg_stats;
! ERROR: locale operation to be invoked, but no collation was derived
-- test variadic polymorphic functions
create function myleast(variadic anyarray) returns anyelement as $$
select min($1[i]) from generate_subscripts($1,1) g(i)

Is there something I can do to help investigate this ?

Regards,

Rémi Zara

On lör, 2011-02-12 at 13:34 +0100, Rémi Zara wrote:
> Since the per-column collation patch went in, pika (NetBSD 5.1/mips) started failing consistently with this diff:
>
> *** /home/pgbuildfarm/workdir/HEAD/pgsql.15101/src/test/regress/expected/polymorphism.out Sat Feb 12 02:16:07 2011
> --- /home/pgbuildfarm/workdir/HEAD/pgsql.15101/src/test/regress/results/polymorphism.out Sat Feb 12 09:10:21 2011
> ***************
> *** 624,630 ****
>
> -- such functions must protect themselves if varying element type isn't OK
> select max(histogram_bounds) from pg_stats;
> ! ERROR: cannot compare arrays of different element types
> -- test variadic polymorphic functions
> create function myleast(variadic anyarray) returns anyelement as $$
> select min($1[i]) from generate_subscripts($1,1) g(i)
> --- 624,630 ----
>
> -- such functions must protect themselves if varying element type isn't OK
> select max(histogram_bounds) from pg_stats;
> ! ERROR: locale operation to be invoked, but no collation was derived
> -- test variadic polymorphic functions
> create function myleast(variadic anyarray) returns anyelement as $$
> select min($1[i]) from generate_subscripts($1,1) g(i)
>
> Is there something I can do to help investigate this ?

It's only failing on this one machine, but there isn't anything
platform-specific in this code, so I'd look for memory management faults
on the code or a compiler problem. Try with lower optimization for a
start.

Le 12 févr. 2011 à 18:51, Peter Eisentraut a écrit :

> On lör, 2011-02-12 at 13:34 +0100, Rémi Zara wrote:
>> Since the per-column collation patch went in, pika (NetBSD 5.1/mips) started failing consistently with this diff:
>>
>> *** /home/pgbuildfarm/workdir/HEAD/pgsql.15101/src/test/regress/expected/polymorphism.out Sat Feb 12 02:16:07 2011
>> --- /home/pgbuildfarm/workdir/HEAD/pgsql.15101/src/test/regress/results/polymorphism.out Sat Feb 12 09:10:21 2011
>> ***************
>> *** 624,630 ****
>>
>> -- such functions must protect themselves if varying element type isn't OK
>> select max(histogram_bounds) from pg_stats;
>> ! ERROR: cannot compare arrays of different element types
>> -- test variadic polymorphic functions
>> create function myleast(variadic anyarray) returns anyelement as $$
>> select min($1[i]) from generate_subscripts($1,1) g(i)
>> --- 624,630 ----
>>
>> -- such functions must protect themselves if varying element type isn't OK
>> select max(histogram_bounds) from pg_stats;
>> ! ERROR: locale operation to be invoked, but no collation was derived
>> -- test variadic polymorphic functions
>> create function myleast(variadic anyarray) returns anyelement as $$
>> select min($1[i]) from generate_subscripts($1,1) g(i)
>>
>> Is there something I can do to help investigate this ?
>
> It's only failing on this one machine, but there isn't anything
> platform-specific in this code, so I'd look for memory management faults
> on the code or a compiler problem. Try with lower optimization for a
> start.
>

Same failure with -O0 (and more shared memory).

Regards,

Rémi Zara

Le 14 févr. 2011 à 19:27, Rémi Zara a écrit :

>
> Le 12 févr. 2011 à 18:51, Peter Eisentraut a écrit :
>
>>
>> It's only failing on this one machine, but there isn't anything
>> platform-specific in this code, so I'd look for memory management faults
>> on the code or a compiler problem. Try with lower optimization for a
>> start.
>>
>
>
> Same failure with -O0 (and more shared memory).
>

Hi,

Without me changing anything (still at -O0), the same error occurred but at the installCheck step, rather than the check step (which passed)

Anything else to try ?

Regards,

Rémi Zara

Le 14 févr. 2011 à 19:27, Rémi Zara a écrit :

>
> Le 12 févr. 2011 à 18:51, Peter Eisentraut a écrit :
>
>>
>> It's only failing on this one machine, but there isn't anything
>> platform-specific in this code, so I'd look for memory management faults
>> on the code or a compiler problem. Try with lower optimization for a
>> start.
>>
>
>
> Same failure with -O0 (and more shared memory).
>

Hi,

Without me changing anything (still at -O0), the same error occurred but at the installCheck step, rather than the check step (which passed)

Anything else to try ?

Regards,

Rémi Zara

=?iso-8859-1?Q?R=E9mi_Zara?= <remi_zara(at)mac(dot)com> writes:
> Le 12 fvr. 2011 18:51, Peter Eisentraut a crit :
>> It's only failing on this one machine, but there isn't anything
>> platform-specific in this code, so I'd look for memory management faults
>> on the code or a compiler problem. Try with lower optimization for a
>> start.

> Same failure with -O0 (and more shared memory).

Note that the query that is failing is an intentional probe of
robustness:

-- check that we can apply functions taking ANYARRAY to pg_stats ...
-- such functions must protect themselves if varying element type isn't OK
select max(histogram_bounds) from pg_stats;
ERROR: cannot compare arrays of different element types

pika is instead showing

ERROR: locale operation to be invoked, but no collation was derived

which some trawling through the code says is coming from varstr_cmp
when fn_collation didn't get set on the call.

Hypothesis: the platform-dependency here is just one of row ordering,
to wit, on most platforms the scan of pg_statistic fails before it gets
to the case where the collation issue is triggered. I'm suspicious that
if two text arrays get compared via this code path, fn_collation fails
to get set, and it's not a platform-specific omission.

It'd be helpful if you could identify the specific values that are
getting compared at the moment of the failure.

regards, tom lane

Le 18 févr. 2011 à 08:26, Tom Lane a écrit :

> =?iso-8859-1?Q?R=E9mi_Zara?= <remi_zara(at)mac(dot)com> writes:
>> Le 12 févr. 2011 à 18:51, Peter Eisentraut a écrit :
>>> It's only failing on this one machine, but there isn't anything
>>> platform-specific in this code, so I'd look for memory management faults
>>> on the code or a compiler problem. Try with lower optimization for a
>>> start.
>
>> Same failure with -O0 (and more shared memory).
>
> Note that the query that is failing is an intentional probe of
> robustness:
>
> -- check that we can apply functions taking ANYARRAY to pg_stats ...
> -- such functions must protect themselves if varying element type isn't OK
> select max(histogram_bounds) from pg_stats;
> ERROR: cannot compare arrays of different element types
>
> pika is instead showing
>
> ERROR: locale operation to be invoked, but no collation was derived
>
> which some trawling through the code says is coming from varstr_cmp
> when fn_collation didn't get set on the call.
>
> Hypothesis: the platform-dependency here is just one of row ordering,
> to wit, on most platforms the scan of pg_statistic fails before it gets
> to the case where the collation issue is triggered. I'm suspicious that
> if two text arrays get compared via this code path, fn_collation fails
> to get set, and it's not a platform-specific omission.
>
> It'd be helpful if you could identify the specific values that are
> getting compared at the moment of the failure.
>

Hi,

Here is what I get after adding an elog in varstr_cmp:

WARNING: Comparing "B011 " and "<fetch first clause> in subqueries^?^?\xa0"
ERROR: locale operation to be invoked, but no collation was derived
STATEMENT: select max(histogram_bounds) from pg_stats;

Regards,

Rémi Zara