Re: logical decoding and replication of sequences

Hi,

Attached is a rebased patch, addressing most of the remaining issues.
The main improvements are:


1) pg_publication_namespace.pntype and type checks

Originally, the patch used pnsequences flag to distinguish which entries
added by FOR ALL TABLES IN SCHEMA and FOR ALL SEQUENCES IN SCHEMA. I've
decided to replace this with a simple char column, called pntype, where
't' means tables and 's' sequences. As explained before, relkind doesn't
work well because of partitioned tables. A char, with a function to
match it to relkind values works fairly well.

I've revisited the question how to represent publications publishing the
same schema twice - once for tables, once for sequences. There were
proposals to represent this with a single row, i.e. turn pntype into an
array of char values. So it'd be either ['t'], ['s'] or ['s', 't']. I
spent some time working on that, but I've decided to keep the current
approach with two separate rows - it's easier to manage, lookup etc.


2) pg_get_object_address

I've updated the objectaddress code to consider pntype when looking-up
the pntype value, so each row in pg_publication_namespace gets the
correct ObjectAddress.


3) for all [tables | sequences]

The original patch did not allow creating publication for all tables and
all sequences at the same time. I've tweaked the grammar to allow this:

   CREATE PUBLICATION p FOR ALL list_of_types;

where list_of_types is arbitrary combination of TABLES and SEQUENCES.
It's implemented in a slightly awkward way - partially in the grammar,
partially in the publicationcmds.c. I suspect there's a (cleaner) way to
do this entirely in the grammar but I haven't succeeded yet.


4) prevent 'ADD TABLE sequence' and 'ADD SEQUENCE table'

It was possible to do "ADD TABLE" and pass it a sequence, which would
fail to notice if the publication already includes all sequences from
the schema. I've added a check preventing that (and a similar one for
ADD SEQUENCE).


5) missing block in AlterTableNamespace to cross-check moving published
sequence to already published schema

A block of code was missing from AlterTableNamespace, checking that
we're not moving a sequence into a schema that is already published (all
the sequences from it).


6) a couple comment fixes

Various comment improvements and fixes. At this point there's a couple
trivial FIXME/XXX comments remaining.


regards

-- 
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company