Re: Sequence Access Method WIP

Hi,

here is another try at this. I had some more time to think about various 
corner cases (like the sequence not having TOAST).

I decided for this version to try bit different approach and that is to 
use custom datatype for the sequence access method state. So there is 
one extra amstate column and the access method defines the custom type.

The idea is that the custom type should be fixed width so that we can do 
proper checks during the creation of access method for size and stuff. 
It also makes the interface bit cleaner as we can provide method for 
reading and saving of the state and abstract the internals more nicely. 
It also solves the issues with critical sections vs memory allocations 
when the access methods wants to save completely new state.

It also means that get_state and set_state interfaces just work with 
text representation of custom type, no need to have complex conversions 
to arrays and back.

One downside is that this means access methods can't use composite types 
as those are not fixed width (nor plain) so the names of the individual 
items in the output are somewhat hidden (see the modified create_view 
test for example of how this affects things).

Other than that, this is based on the new am api by Alexander Korotkov 
[1]. It extends it by adding another column called amkind to the pg_am 
which can have either value "i" for index or "S" for sequence (same as 
relkind in pg_class for those).

I didn't attach DDL and the gapless sequence yet, mainly because I don't 
want to waste time rewriting it again in case we can't agree that this 
approach is good (based on the long discussion and resulting several 
rewrites wrt the multiple columns vs bytea previously).

[1] https://commitfest.postgresql.org/8/336/

-- 
  Petr Jelinek                  http://www.2ndQuadrant.com/
  PostgreSQL Development, 24x7 Support, Training & Services