On 10/27/2014 03:46 PM, Tom Lane wrote:
> Heikki Linnakangas <hlinnakangas@vmware.com> writes:
>> On 10/27/2014 03:21 PM, Tomas Vondra wrote:
>>> Thinking about this a bit more, do we really need a full checkpoint? That
>>> is a checkpoint of all the databases in the cluster? Why checkpointing the
>>> source database is not enough?
>
>> A full checkpoint ensures that you always begin recovery *after* the
>> DBASE_CREATE record. I.e. you never replay a DBASE_CREATE record during
>> crash recovery (except when you crash before the transaction commits, in
>> which case it doesn't matter if the new database's directory is borked).
>
> Yeah. After re-reading the 2005 thread, I wonder if we shouldn't just
> bite the bullet and redesign CREATE DATABASE as you suggest, ie, WAL-log
> all the copied files instead of doing a "cp -r"-equivalent directory copy.
> That would fix a number of existing replay hazards as well as making it
> safe to do what Tomas wants. In the small scale this would cause more I/O
> (2 copies of the template database's data) but in production situations
> we might well come out ahead by avoiding a forced checkpoint of the rest
> of the cluster. Also I guess we could skip WAL-logging if WAL archiving
> is off, similarly to the existing optimization for CREATE INDEX etc.
That would be a nasty surprise for anyone who's using CREATE DATABASE as
a fast way to clone a large database. But I would be OK with that, at
least if we can skip the WAL-logging with wal_level=minimal.
- Heikki