Re: tableam vs. TOAST

On Wed, Nov 6, 2019 at 12:00 PM Andres Freund <andres@anarazel.de> wrote:
> I'd like an AM to have the *option* of implementing something better, or
> at least go in the direction of making that possible.

OK. Could you see what you think of the attached patches? 0001 does
some refactoring of toast_fetch_datum() and toast_fetch_datum_slice()
to make them look more like each other and clean up a bunch of stuff
that I thought was annoying, and 0002 then pulls out the common logic
into a heap-specific function. If you like this direction, we could
then push the heap-specific function below tableam, but I haven't done
that yet.

> It seems perfectly possible to have a helper function implementing the
> current logic that you just can call with the fixed chunk size as an
> additional parameter. Which'd basically mean there's no meaningful
> difference in complexity compared to providing the chunk size as an
> external AM property. In one case you have a callback that just calls a
> helper function with one parameter, in the other you fill in a member of
> the struct.

I haven't tried to do this yet. I think that to make it work, the
helper function would have to operate in terms of slots instead of
using fastgetattr() as this logic does now. I don't know whether that
would be faster (because the current code might have a little less in
terms of indirect function calls) or slower (because the current code
makes two calls to fastgetattr and if we used slots here we could just
deform once). I suspect it might be a small enough difference not to
worry too much about it either way.

-- 
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company