On Fri, Dec 10, 2010 at 10:33 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> I wrote:
>>> Basically, I'm thinking that given CREATE TABLESPACE LOCATION '/foo/bar'
>>> the creation and properties of /foo/bar/PG_9.0_201004261 ought to be
>>> handled *exactly* the way that the -D target directory of initdb is.
>
> One interesting point here is that initdb uses the equivalent of mkdir
> -p, so it will automatically try to create parent directories of
> whatever path you specify. Duplicating that behavior in CREATE
> TABLESPACE causes this diff in the regression tests:
>
> -- Will fail with bad path
> CREATE TABLESPACE badspace LOCATION '/no/such/location';
> ! ERROR: directory "/no/such/location" does not exist
> -- No such tablespace
> CREATE TABLE bar (i int) TABLESPACE nosuchspace;
> ERROR: tablespace "nosuchspace" does not exist
> --- 65,71 ----
>
> -- Will fail with bad path
> CREATE TABLESPACE badspace LOCATION '/no/such/location';
> ! ERROR: could not create directory "/no": Permission denied
> -- No such tablespace
> CREATE TABLE bar (i int) TABLESPACE nosuchspace;
> ERROR: tablespace "nosuchspace" does not exist
>
> I'm not sure that this is a bad thing. In particular, it makes WAL
> replay noticeably more robust since it will do what it can to regenerate
> the whole path if you deleted parent directories. It will of course
> still fail, as here, if the server doesn't have write permissions on the
> last existing dir in the path.
>
> Anybody have a problem with adopting this behavior?
Seems a bit surprising.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company