"Eric G. Miller" <egm2@jps.net> writes:
> While I don't see such dramatic results with this bad input data, I did
> run across something a little peculiar:
With what PG version? On what platform?
Comparing 7.1.1 against 7.0.*, I note that the resolution of "illegal"
times during a DST jump has changed, at least on my platform (HPUX 10.20).
7.0.* resolves the time backwards whereas current sources resolve
forwards: "select timestamp('2001-04-01 02:02:02')" yields
2001-04-01 01:02:02-05 in 7.0.2
2001-04-01 03:02:02-04 in current
Since this is dependent on how the local mktime() library routine
reacts to "illegal" times, some platform-to-platform variation is to be
expected. Your example looks like mktime() must actually have some
internal state on your machine, causing its result to depend on what
it was asked previously :-(
regards, tom lane