On Sat, Apr 24, 2010 at 11:02 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> Pushing it into the RelFileNode has some advantages in terms of being
>> able to get at the information from everywhere, but one thing that
>> makes me think that's probably not a good decision is that we somtimes
>> WAL-log relfilenodes. And WAL-logging the value of the isTemp flag is
>> a waste, because if we're WAL-logging, it's zero.
>
> Yeah. I think we also use RelFileNode as a hash tag in places, and
> so adding a bool to it would be problematic for a couple of reasons:
> possibly uninitialized pad bytes, and uselessly incorporating more bytes
> into the hash calculation.
Right. I was thinking about the padding issue, too. I took a crack
at adding an isTemp argument to smgropen() and removing it everywhere
else, but it turns out this isn't as straightforward as it ought to be
because nbtsort.c, rewriteheap.c, and tablecmds.c feel entitled to
violate the abstraction layer by passing isTemp = true to smgrwrite()
and/or smgrextend() even when the real value is false, under the
assumption that the only thing isTemp is doing in those functions is
controlling fsync behavior. I think we'll have to replace the isTemp
argument to those functions with a boolean whose explicit charter is
to do what they're using it for.
Removing the isTemp from smgrtruncate() and smgrdounlink() argument
looks easy, though. WIP patch that takes it that far is attached.
[davidfetter: patch description for PWN is "WIP patch to push isTemp
down into the smgr layer"]
...Robert
