Oleg Bartunov <oleg@sai.msu.su> writes:
>> What PG version is this exactly?
> REL8_0_STABLE, I believe. I posted another problem, now "cut and pasted".
I've been able to duplicate this here. What is happening is that the
damage to ./t1 is being done when you start the postmaster in ./t2.
It looks to me like the fundamental problem is that the t2 postmaster
is replaying the WAL-logged CREATE DATABASE command from t1's xlog, and
*that log entry contains an absolute path name*. So the CREATE replay
is wiping out t1's "test" database subdirectory.
This isn't a problem in normal use of course, but it'd be a serious
issue for someone engaging in WAL-shipping, if their backup postmaster
were living at a different absolute path. We probably need to think
about whether we can make CREATE DATABASE log only relative paths.
Log-shipping CREATE TABLESPACE commands is even more interesting :-(.
Not sure how to deal with that.
regards, tom lane