Alvaro Herrera <alvherre@commandprompt.com> writes:
> Maybe we could make this work by fiddling with a different smgr -- on
> it, smgr_sync would be a noop, as would smgr_immedsync, and we could
> kludge something up to truncate relations during recovery.
Interesting thought but I think it falls down on pg_statistic.
One comment I've got is that we have already concluded that the spec's
GLOBAL/LOCAL TEMP TABLE distinction is not related to cross-session
persistence of the table definitions, but rather to module visibility
which is a concept we have not got (yet). Ergo, we should not use the
phrase "global temp table" for these things. Not sure what to suggest
instead. Perhaps call them "session tables" instead of "temp tables"?
regards, tom lane