Hi,
On 2024-04-16 08:31:24 -0400, Robert Haas wrote:
> On Tue, Apr 16, 2024 at 6:52 AM Alexander Korotkov <aekorotkov@gmail.com> wrote:
> > Taking a closer look at acquire_sample_rows(), I think it would be
> > good if table AM implementation would care about block-level (or
> > whatever-level) sampling. So that acquire_sample_rows() just fetches
> > tuples one-by-one from table AM implementation without any care about
> > blocks. Possible table_beginscan_analyze() could take an argument of
> > target number of tuples, then those tuples are just fetches with
> > table_scan_analyze_next_tuple(). What do you think?
>
> Andres is the expert here, but FWIW, that plan seems reasonable to me.
> One downside is that every block-based tableam is going to end up with
> a very similar implementation, which is kind of something I don't like
> about the tableam API in general: if you want to make something that
> is basically heap plus a little bit of special sauce, you have to copy
> a mountain of code. Right now we don't really care about that problem,
> because we don't have any other tableams in core, but if we ever do, I
> think we're going to find ourselves very unhappy with that aspect of
> things. But maybe now is not the time to start worrying. That problem
> isn't unique to analyze, and giving out-of-core tableams the
> flexibility to do what they want is better than not.
I think that can partially be addressed by having more "block oriented AM"
helpers in core, like we have for table_block_parallelscan*. Doesn't work for
everything, but should for something like analyze.
Greetings,
Andres Freund