On Thu, Apr 9, 2015 at 5:52 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
> On 9 April 2015 at 04:12, Michael Paquier <michael.paquier@gmail.com> wrote:
>
>> Also, I am wondering if the sampling logic based on block analysis is
>> actually correct, for example for now this fails and I think that we
>> should support it:
>> =# with query_select as (select generate_series(1, 10) as a) select
>> query_select.a from query_select tablesample system (100.0/11)
>> REPEATABLE (9999);
>> ERROR: 42P01: relation "query_select" does not exist
>>
>> How does the SQL spec define exactly TABLESAMPLE? Shouldn't we get a
>> sample from a result set?
>> Thoughts?
>
> Good catch. The above query doesn't make any sense.
>
> TABLESAMPLE SYSTEM implies system-defined sampling mechanism, which is
> block level sampling. So any block level sampling method should be
> barred from operating on a result set in this way... i.e. should
> generate an "ERROR inappropriate sampling method specified"
>
> TABLESAMPLE BERNOULLI could work in this case, or any other non-block
> based sampling mechanism. Whether it does work yet is another matter.
>
> This query should be part of the test suite and should generate a
> useful message or work correctly.
Ahah, you just beat me on that ;) See a more precise reply below.
--
Michael