On Sat, Mar 4, 2017 at 11:41 AM, Petr Jelinek
<petr.jelinek@2ndquadrant.com> wrote:
> On 04/03/17 06:46, Robert Haas wrote:
>> On Tue, Feb 28, 2017 at 12:08 PM, Erik Rijkers <er@xs4all.nl> wrote:
>>> Would you remind me why synchronous_commit = on was deemed a better default?
>> I'm wondering about that, too. If you're trying to do logical
>> synchronous replication, then maybe there's some argument there,
>> although even in that case I am not sure it's actually necessary. But
>> if you're doing asynchronous logical replication, it seems not to make
>> much sense. I mean, walwriter is going to flush the WAL to disk
>> within a fraction of a second; why would we wait for that to happen
>> instead of getting on with replicating the next transaction meanwhile?
>>
>
> I do agree. And even when doing synchronous replication it should not be
> problem. We do have code already in place which takes care of reporting
> correct write/flush position in the situation where upstream does
> syncrep but downstream has synccommit off.
>
> The only anomaly I could think of in terms of having sync commit off on
> downstream is that some replicated data that were visible will disappear
> after a crash only to reappear again once replication catches up to
> where it has been before when doing asynchronous replication. That does
> not strike me like good enough reason to have by default something like
> order of magnitude worse performance.
I see. That's a good reason to have an option, but I agree with your
conclusion.
> So I think we should do it, but it needs to be configurable, my original
> patch added GUC for it, Peter wanted it to be configurable per
> subscription. I guess we could add it as another option to the list of
> WITH (...) options for CREATE and ALTER SUBSCRIPTION.
I don't have a terribly well-considered opinion on this point just
yet, but my initial hunch is that Peter has the right idea.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
