"Simon Riggs" <simon@2ndquadrant.com> writes:
> Good hunch. I plugged this hole earlier, but on further inspection I can
> see the plug wasn't wide enough. XLogAsyncCommitFlush() is good enough,
> but HeapTupleSatisfiesVacuum() still allowed the inexact bookkeeping to
> sometimes skip hint bit setting, when executed with concurrent
> transactions touching other tables.
> ISTM that if we call HeapTupleSatisfiesVacuum() with an additional
> boolean parameter, force, we can tell VF to always set the hint bits in
> every case, not just HEAP_MOVED_IN and HEAP_MOVED_OUT.
Surely this approach is no good: won't it allow hint bits to reach disk
in advance of their transaction?
I think it'd be safer, and a lot less ugly, to recast the tests in
VACUUM FULL. If we make the first pass clear any old MOVED_IN/MOVED_OUT
bits then the last pass can key off those instead of assuming that
XMIN_COMMITTED is set everywhere. Then we'd not need
XLogAsyncCommitFlush, which is a kluge anyway.
regards, tom lane