Github user HyukjinKwon commented on the issue: https://github.com/apache/spark/pull/19234 Hm, BTW, do we handle https://github.com/python/cpython/blob/018d353c1c8c87767d2335cd884017c2ce12e045/Lib/datetime.py#L1443-L1455: ```python if tz is None: # As of version 2015f max fold in IANA database is # 23 hours at 1969-09-30 13:00:00 in Kwajalein. # Let's probe 24 hours in the past to detect a transition: max_fold_seconds = 24 * 3600 y, m, d, hh, mm, ss = converter(t - max_fold_seconds)[:6] probe1 = cls(y, m, d, hh, mm, ss, us, tz) trans = result - probe1 - timedelta(0, max_fold_seconds) if trans.days < 0: y, m, d, hh, mm, ss = converter(t + trans // timedelta(0, 1))[:6] probe2 = cls(y, m, d, hh, mm, ss, us, tz) if probe2 == result: result._fold = 1 ``` Or do you guys see it could be ignorable as it is quite newly fixed (in 3.6.x) and a corner case vs the improvement? Looking at the performance improvement in the PR description, it sounds pretty trivial. If it is safe to go, I am okay but if we miss anything, I doubt if it is worth fixing.
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org