Re: changeset generation v5-01 - Patches & git tree

On 2013-07-01 14:16:55 -0400, Tom Lane wrote:
> Alvaro Herrera <alvherre(at)2ndquadrant(dot)com> writes:
> > So the question is, do we take the overhead of the new index (which
> > means overhead on DML operations -- supposedly rare) or do we take the
> > overhead of larger WAL records (which means overhead on all DDL
> > operations)?
>
> > Note we can make either thing apply to only people running logical
> > replication.
>
> I don't believe you can have or not have an index on pg_class as easily
> as all that. The choice would have to be frozen at initdb time, so
> people would have to pay the overhead if they thought there was even a
> small possibility that they'd want logical replication later.

It should be possible to create the index in a single database when we
start logical replication in that database? Running the index creation
with a fixed oid shouldn't require too much code. The oid won't be
reused by other pg_class entries since it would be a system one.
Alternatively we could always create the index's pg_class/index entry
but mark it as !indislive when logical replication isn't active for that
database. Then activating it would just require rebuilding that
index.

But then, I am not fully convinced that's worth the trouble since I
don't think pg_class index maintenance is the painspot in DDL atm.

> Flipping the content of WAL records might not be a terribly simple thing
> to do either, but at least in principle it could be done during a
> postmaster restart, without initdb.

The main patch combines various booleans in the heap wal records into a
flags variable, so there should be enough space to keep track of it
without increasing size. Makes size calculations a bit more annoying
though as we use the xlog record length to calculate the heap tuple's
length, but that's not a large problem.
So we could just set the XLOG_HEAP_CONTAINS_CLASSOID flag if wal_level
>= WAL_LEVEL_LOGICAL. Wal decoding then can throw a tantrum if it finds
a record without it and we're done.

We could even make that per database, but that seems to be something for
the future.

Greetings,

Andres Freund

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