The underlying issue (but not strictly a bug as the documentation
specifically says not to do that -
http://sources.debian.net/src/python-tz/2016.7-0.2/pytz/tzinfo.py/#L247
) is that passing a pytz tzinfo to the datetime constructor uses its
first listed offset, not its correct offset for that date:
>>> datetime.datetime(2017,4,1,tzinfo=pytz.timezone('Europe/London'))
datetime.datetime(2017, 4, 1, 0, 0, tzinfo=<DstTzInfo 'Europe/London'
GMT0:00:00 STD>)
>>> pytz.timezone('Europe/London').localize(datetime.datetime(2017,4,1))
datetime.datetime(2017, 4, 1, 0, 0, tzinfo=<DstTzInfo 'Europe/London'
BST+1:00:00 DST>)
This suggests the attached fix, but this has *not* been tested.
As for why it's only now showing up...there seems to be something weird
going on where the initial 'LMT' entries of timezone definitions
sometimes get loaded into pytz.timezone objects but often don't:
jessie>>> pytz.timezone('Asia/Tokyo')._tzinfos
{(datetime.timedelta(0, 32400), datetime.timedelta(0), 'JST'):
<DstTzInfo 'Asia/Tokyo' JST+9:00:00 STD>, (datetime.timedelta(0, 32400),
datetime.timedelta(0), 'JCST'): <DstTzInfo 'Asia/Tokyo' JCST+9:00:00
STD>, (datetime.timedelta(0, 36000), datetime.timedelta(0, 3600),
'JDT'): <DstTzInfo 'Asia/Tokyo' JDT+10:00:00 DST>}
sid>>> pytz.timezone('Asia/Tokyo')._tzinfos
{(datetime.timedelta(0, 32400), datetime.timedelta(0), 'JST'):
<DstTzInfo 'Asia/Tokyo' JST+9:00:00 STD>, (datetime.timedelta(0, 36000),
datetime.timedelta(0, 3600), 'JDT'): <DstTzInfo 'Asia/Tokyo'
JDT+10:00:00 DST>, (datetime.timedelta(0, 33540), datetime.timedelta(0),
'LMT'): <DstTzInfo 'Asia/Tokyo' LMT+9:19:00 STD>}
--- pandas/tests/test_multilevel0.py 2017-04-01 23:02:44.659970299 +0100
+++ pandas/tests/test_multilevel.py 2017-04-01 22:55:29.031975195 +0100
@@ -84,9 +84,9 @@ class TestMultiLevel(tm.TestCase):
# GH 7112
import pytz
tz = pytz.timezone('Asia/Tokyo')
- expected_tuples = [(1.1, datetime.datetime(2011, 1, 1, tzinfo=tz)),
- (1.2, datetime.datetime(2011, 1, 2, tzinfo=tz)),
- (1.3, datetime.datetime(2011, 1, 3, tzinfo=tz))]
+ expected_tuples = [(1.1, tz.localize(datetime.datetime(2011, 1, 1))),
+ (1.2, tz.localize(datetime.datetime(2011, 1, 2))),
+ (1.3, tz.localize(datetime.datetime(2011, 1, 3)))]
expected = Index([1.1, 1.2, 1.3] + expected_tuples)
self.assertTrue(result.equals(expected))
@@ -104,9 +104,9 @@ class TestMultiLevel(tm.TestCase):
result = midx_lv3.append(midx_lv2)
expected = Index._simple_new(
- np.array([(1.1, datetime.datetime(2011, 1, 1, tzinfo=tz), 'A'),
- (1.2, datetime.datetime(2011, 1, 2, tzinfo=tz), 'B'),
- (1.3, datetime.datetime(2011, 1, 3, tzinfo=tz), 'C')] +
+ np.array([(1.1, tz.localize(datetime.datetime(2011, 1, 1)), 'A'),
+ (1.2, tz.localize(datetime.datetime(2011, 1, 2)), 'B'),
+ (1.3, tz.localize(datetime.datetime(2011, 1, 3)), 'C')] +
expected_tuples), None)
self.assertTrue(result.equals(expected))
