2008/12/25 Pavel Stehule <pavel.stehule@gmail.com>:
> 2008/12/25 Hitoshi Harada <umi.tanuki@gmail.com>:
>> 2008/12/25 Greg Stark <greg.stark@enterprisedb.com>:
>>> Yeah, it seems like adding a flag like iswindowable to aggregate functions
>>> is the safest option.
>>>
>>> It would be nice if it represented an abstract property of the state
>>> function or final function rather than just "works with the implementation
>>> of window functions". I'm not sure what that property is though -
>>> isidempotent? isreentrant? Maybe just a vague isrepeatable?
>>
>> No, I meant wrinting such like:
>>
>> Datum
>> some_trans_fn(PG_FUNCTION_ARGS)
>> {
>> if (fcinfo->context && IsA(fcinfo->context, WindowAggState))
>> elog(ERROR, "some_agg does not support window aggregate");
>>
>> ...
>> }
>>
>> rather than adding column to catalog. To add flag you must add new
>> syntax for CREATE AGGREGATE, which is slightly more painful.
>>
>
> enhancing of CREATE AGGREGATE syntax should be better, it could solve
> problem with compatibility.
>
Most of the aggregates have no problem with this issue and the rest
are fixable like array_agg. So adding a column and syntax is too much,
I guess. My suggestion of raising error is urgent patch for third
party aggregates that are copied from contrib/intagg.
But if there is a chance of general approach to call repeatable
aggregate other than WindowAgg, that should be considered.
Regards,
--
Hitoshi Harada