Michael Fuhr <mike@fuhr.org> writes:
> I can duplicate this in 7.3.10, but only if the postmaster does a
> redo when it restarts. If I do a checkpoint before killing the
> postmaster then the insert succeeds.
> I couldn't duplicate this behavior in 7.4.8, 8.0.3, or HEAD. The
> 7.4 Release Notes have an item about making B-tree indexes fully
> WAL-safe, so I wonder if that fixes the problem.
Yeah. Before 7.4 there was no WAL record emitted for the act of
initializing the btree metapage, so this behavior is pretty much
exactly what you'd expect. It's possible that we could backport
that single change, but if memory serves there were a number of
other ways in which btree index build violated the WAL principle,
so I'm not sure there's much point. You more or less had to
checkpoint to be sure the new index is fully down to disk.
regards, tom lane