On 01/03/2017 01:34 PM, Michael Paquier wrote:
> On Mon, Jan 2, 2017 at 10:55 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
>> In the hope of making things better in 10.0, I remove my objection.
>> If people want to use wal_level = minimal they can restart their
>> server and they can find that out in the release notes.
>>
>> Should we set wal_level = replica or wal_level = logical as the
>> default for 10.0?
>
> replica sounds like a better default to me as most users use at
> least archiving. Logical decoding is still fresh though, and its use
> is not that wide. Have there been any study on its performance
> impact compared to replica by the way?
>
I've just posted results for some benchmarks I'm running. Those are some
simple pgbench tests, nothing special, and according to those results
the performance impact (logical vs. replica) is negligible. I can run
additional tests with other workloads, of course.
While we can probably construct workloads where the difference is
measurable, I'm not sure performance impact is the main concern here. As
you point out, we have 'replica' since 9.0 effectively, while logical is
much newer, so perhaps there are some hidden bugs? It'd be embarrassing
to pick 'logical' and hurt everyone, even if they don't get any benefit
from wal_level=logical.
So +1 to 'replica' and allowing switching to 'logical' without restart.
That should not be extremely difficult, as the main culprit seems to be
max_wal_senders/max_replication_slots requiring shared memory. But with
'replica' we already have those enabled/allocated, unlike when switching
from 'minimal' to 'replica'.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services