On Thu, May 3, 2012 at 2:23 AM, Simon Riggs <simon@2ndquadrant.com> wrote:
> On Wed, May 2, 2012 at 9:38 PM, Daniel Farina <daniel@heroku.com> wrote:
>
>> Besides accuracy, there is a thornier problem here that has to do with
>> hot standby (although the use case is replication more generally) when
>> one has heterogeneously sized database resources. As-is, it is
>> required that locking-related structures -- max_connections,
>> max_prepared_xacts, and max_locks_per_xact (but not predicate locks,
>> is that an oversight?) must be a larger number on a standby than on a
>> primary.
>
>>= not >
> so you can use the same values on both sides
>
> Predicate locks aren't set in recovery so the value isn't checked as a
> required parameter value.
I had a feeling that might be the case, since my understanding is that
they are not actually locks -- rather, markers.
In any case, it would be strange to change the *number* of locks per
transaction in such heterogeneous environments because then some
fairly modestly sized transactions will simply not work depending on
one size of system one selects. The more problematic issue is that
small systems will be coerced into having a very high number for
max_connections and the memory usage required by that, if one also
provides a large system supporting a high connection limit and moves
things around via WAL shipping. I'm not sure what there is to be done
about this other than make the absolutely required locking structures
smaller -- I wonder if not unlike the out-of-line storage for PGPROC
patch this might also make some things faster. All in all, without
having gone in to figure out *why* the size consumption is as it is
I'm a little flabbergasted as to why the locking structures are just
so large.
--
fdr
