Petr Jelinek <petr@2ndquadrant.com> writes:
> On 2015-07-12 18:02, Tom Lane wrote:
>> A possible way around this problem is to redefine the sampling rule so
>> that it is not history-dependent but depends only on the tuple TIDs.
>> For instance, one could hash the TID of a candidate tuple, xor that with
>> a hash of the seed being used for the current query, and then select the
>> tuple if (hash/MAXINT) < P.
> That would work for bernoulli for physical tuples, yes. Only thing that
> worries me is future extensibility for data sources that only provide
> virtual tuples.
Well, repeatability of a TABLESAMPLE attached to a join seems like an
unsolved and possibly unsolvable problem anyway. I don't think we should
assume that the API we define today will cope with that.
But that is another reason why the current API is inadequate: there's no
provision for specifying whether or how a tablesample method can be
applied to non-base-table RTEs. (I re-read the thread and noted that
Peter E. complained about that some time ago, but nothing was done about
it. I'm fine with not supporting the case right now, but nonetheless
it's another reason why we'd better make the API more easily extensible.)
regards, tom lane