On 11/4/21 23:56, Tomas Vondra wrote:
> Hi,
>
> On 11/4/21 17:53, Justin Pryzby wrote:
>> On Thu, Nov 04, 2021 at 09:46:49AM +0100, Andreas Seltenreich wrote:
>>> sqlsmith triggers the following assertion when testing REL_14_STABLE:
>>>
>>> TRAP: FailedAssertion("a1 <= a2", File: "brin_minmax_multi.c",
>>> Line: 1879, PID: 631814)
>>>
>>> I can reproduce it with the following query on a fresh regression
>>> database:
>>>
>>> insert into public.brintest_multi (float4col) values (real 'nan');
>>>
>>> The branch was at f6162c020c while testing, backtrace below.
>>
>> I couldn't reproduce this, but it reminds me of this one, which we
>> also had
>> trouble reproducing.
>>
>
> I can reproduce that just fine - all I had to do was 'make installcheck'
> and then connect to the regression db and run the insert.
>
> It seems to be a simple case of confusion in handling of NaN values. We
> do sort them correctly (by calling float4_lt), but given two values
>
> arg1 = nan (0x400000)
> arg2 = 0.0909090936
>
> then a simple comparison does not give the expected result
>
> (arg1 < arg2)
> (arg1 == arg2)
> (arg1 > arg2)
>
> all evaluate to false, which is why the assert fails. So I guess the
> distance function for float4 (and probably float8 too) need a couple
> more lines checking NaN.
>
Here's a patch that should fix this. It simply handled NaN and returns
distance 0.0 for two NaN values and Infinity for NaN vs. something else.
I did check what happens with Infinity values, but I think those are
fine - we can actually calculate distance with them, and the assert
works for them too.
I'll improve the regression tests to include a couple NaN/Infinity cases
tomorrow, and then I'll get this pushed.
regards
--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
