Stephen Frost <sfrost@snowman.net> writes:
> Tom,
> * Tom Lane (tgl@sss.pgh.pa.us) wrote:
>> Don't see what. The main reason we've not yet attempted a global fix is
>> that the most straightforward way (take a new snapshot each time we
>> start a new SnapshotNow scan) seems too expensive. But CREATE DATABASE
>> is so expensive that the cost of an extra snapshot there ain't gonna
>> matter.
> Patch attached. Passes the regression tests and our internal testing
> shows that it eliminates the error we were seeing (and doesn't cause
> blocking, which is even better).
I committed this with a couple of changes:
* I used GetLatestSnapshot rather than GetTransactionSnapshot. Since
we don't allow these operations inside transaction blocks, there
shouldn't be much difference, but in principle GetTransactionSnapshot
is the wrong thing; in a serializable transaction it could be quite old.
* After reviewing the other uses of SnapshotNow in dbcommands.c, I
decided we'd better make the same type of change in remove_dbtablespaces
and check_db_file_conflict, because those are likewise doing filesystem
operations on the strength of what they find in pg_tablespace.
I also ended up deciding to back-patch to 8.3 as well.
regards, tom lane