Alvaro Herrera <alvherre@commandprompt.com> writes:
> Okay, here's a patch implementing this idea. It seems to work quite
> well, and it solves the problem in a limited testing scenario -- I
> haven't yet tested on the customer machines.
This seems mostly sane, except I think you have not considered the
issue of when to clear the smgr_transient flag on an existing
SMgrRelation: if it starts getting used for "normal" accesses after
having by chance been used for a blind write, we don't want the
transient marking to persist. That's why I suggested having smgropen
always clear it.
Likewise, I think the FD_XACT_TRANSIENT flag on a VFD needs to go away
at some point, probably once it's actually been closed at EOXACT, though
there's doubtless more than one way to handle that.
> This customer is running on 8.4 so I started from there; should I
> backpatch this to 8.2, or not at all?
I'm not excited about back-patching it...
regards, tom lane