matplotlib · timhoffm · Apr 29, 2020 · Aug 7, 2019
diff --git a/doc/api/api_changes_3.3/deprecations.rst b/doc/api/api_changes_3.3/deprecations.rst
@@ -491,3 +491,8 @@ experimental and may change in the future.
 ``testing.compare.make_external_conversion_command``
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 ... is deprecated.
+
+`.epoch2num` and `.num2epoch` are deprecated
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+These are unused and can be easily reproduced by other date tools.
+`.get_epoch` will return Matplotlib's epoch.
diff --git a/doc/users/next_whats_new/2020-04-09-datetime-epoch-change.rst b/doc/users/next_whats_new/2020-04-09-datetime-epoch-change.rst
@@ -0,0 +1,22 @@
+Dates now use a modern epoch
+----------------------------
+
+Matplotlib converts dates to days since an epoch using `.dates.date2num` (via
+`matplotlib.units`).  Previously, an epoch of ``0000-12-31T00:00:00`` was used
+so that ``0001-01-01`` was converted to 1.0.  An epoch so distant in the
+past meant that a modern date was not able to preserve microseconds because
+2000 years times the 2^(-52) resolution of a 64-bit float gives 14
+microseconds.
+
+Here we change the default epoch to the more reasonable UNIX default of
+``1970-01-01T00:00:00`` which for a modern date has 0.35 microsecond
+resolution.  (Finer resolution is not possible because we rely on
+`datetime.datetime` for the date locators). Access to the epoch is provided
+by `~.dates.get_epoch`, and there is a new :rc:`date.epoch` rcParam.  The user
+may also call `~.dates.set_epoch`, but it must be set *before* any date
+conversion or plotting is used.
+
+If you have data stored as ordinal floats in the old epoch, a simple
+conversion (using the new epoch) is::
+
+    new_ordinal = old_ordinal + mdates.date2num(np.datetime64('0000-12-31'))
diff --git a/examples/ticks_and_spines/date_precision_and_epochs.py b/examples/ticks_and_spines/date_precision_and_epochs.py
@@ -0,0 +1,155 @@
+"""
+=========================
+Date Precision and Epochs
+=========================
+
+Matplotlib can handle `.datetime` objects and `numpy.datetime64` objects using
+a unit converter that recognizes these dates and converts them to floating
+point numbers.
+
+Before Matplotlib 3.3, the default for this conversion returns a float that was
+days since "0000-12-31T00:00:00".  As of Matplotlib 3.3, the default is
+days from "1970-01-01T00:00:00".  This allows more resolution for modern
+dates.  "2020-01-01" with the old epoch converted to 730120, and a 64-bit
+floating point number has a resolution of 2^{-52}, or approximately
+14 microseconds, so microsecond precision was lost.  With the new default
+epoch "2020-01-01" is 10957.0, so the achievable resolution is 0.21
+microseconds.
+
+"""
+import datetime
+import numpy as np
+
+import matplotlib
+import matplotlib.pyplot as plt
+import matplotlib.dates as mdates
+
+
+def _reset_epoch_for_tutorial():
+    """
+    Users (and downstream libraries) should not use the private method of
+    resetting the epoch.
+    """
+    mdates._reset_epoch_test_example()
+
+
+#############################################################################
+# Datetime
+# --------
+#
+# Python `.datetime` objects have microsecond resolution, so with the
+# old default matplotlib dates could not round-trip full-resolution datetime
+# objects.
+
+old_epoch = '0000-12-31T00:00:00'
+new_epoch = '1970-01-01T00:00:00'
+
+_reset_epoch_for_tutorial()  # Don't do this.  Just for this tutorial.
+mdates.set_epoch(old_epoch)  # old epoch (pre MPL 3.3)
+
+date1 = datetime.datetime(2000, 1, 1, 0, 10, 0, 12,
+                          tzinfo=datetime.timezone.utc)
+mdate1 = mdates.date2num(date1)
+print('Before Roundtrip: ', date1, 'Matplotlib date:', mdate1)
+date2 = mdates.num2date(mdate1)
+print('After Roundtrip:  ', date2)
+
+#############################################################################
+# Note this is only a round-off error, and there is no problem for
+# dates closer to the old epoch:
+
+date1 = datetime.datetime(10, 1, 1, 0, 10, 0, 12,
+                          tzinfo=datetime.timezone.utc)
+mdate1 = mdates.date2num(date1)
+print('Before Roundtrip: ', date1, 'Matplotlib date:', mdate1)
+date2 = mdates.num2date(mdate1)
+print('After Roundtrip:  ', date2)
+
+#############################################################################
+# If a user wants to use modern dates at microsecond precision, they
+# can change the epoch using `~.set_epoch`.  However, the epoch has to be
+# set before any date operations to prevent confusion between different
+# epochs. Trying to change the epoch later will raise a `RuntimeError`.
+
+try:
+    mdates.set_epoch(new_epoch)  # this is the new MPL 3.3 default.
+except RuntimeError as e:
+    print('RuntimeError:', str(e))
+
+#############################################################################
+# For this tutorial, we reset the sentinel using a private method, but users
+# should just set the epoch once, if at all.
+
+_reset_epoch_for_tutorial()  # Just being done for this tutorial.
+mdates.set_epoch(new_epoch)
+
+date1 = datetime.datetime(2020, 1, 1, 0, 10, 0, 12,
+                          tzinfo=datetime.timezone.utc)
+mdate1 = mdates.date2num(date1)
+print('Before Roundtrip: ', date1, 'Matplotlib date:', mdate1)
+date2 = mdates.num2date(mdate1)
+print('After Roundtrip:  ', date2)
+
+#############################################################################
+# datetime64
+# ----------
+#
+# `numpy.datetime64` objects have microsecond precision for a much larger
+# timespace than `.datetime` objects.  However, currently Matplotlib time is
+# only converted back to datetime objects, which have microsecond resolution,
+# and years that only span 0000 to 9999.
+
+_reset_epoch_for_tutorial()  # Don't do this.  Just for this tutorial.
+mdates.set_epoch(new_epoch)
+
+date1 = np.datetime64('2000-01-01T00:10:00.000012')
+mdate1 = mdates.date2num(date1)
+print('Before Roundtrip: ', date1, 'Matplotlib date:', mdate1)
+date2 = mdates.num2date(mdate1)
+print('After Roundtrip:  ', date2)
+
+#############################################################################
+# Plotting
+# --------
+#
+# This all of course has an effect on plotting.  With the old default epoch
+# the times were rounded, leading to jumps in the data:
+
+_reset_epoch_for_tutorial()  # Don't do this.  Just for this tutorial.
+mdates.set_epoch(old_epoch)
+
+x = np.arange('2000-01-01T00:00:00.0', '2000-01-01T00:00:00.000100',
+              dtype='datetime64[us]')
+y = np.arange(0, len(x))
+fig, ax = plt.subplots(constrained_layout=True)
+ax.plot(x, y)
+ax.set_title('Epoch: ' + mdates.get_epoch())
+plt.setp(ax.xaxis.get_majorticklabels(), rotation=40)
+plt.show()
+
+#############################################################################
+# For a more recent epoch, the plot is smooth:
+
+_reset_epoch_for_tutorial()  # Don't do this.  Just for this tutorial.
+mdates.set_epoch(new_epoch)
+
+fig, ax = plt.subplots(constrained_layout=True)
+ax.plot(x, y)
+ax.set_title('Epoch: ' + mdates.get_epoch())
+plt.setp(ax.xaxis.get_majorticklabels(), rotation=40)
+plt.show()
+
+_reset_epoch_for_tutorial()  # Don't do this.  Just for this tutorial.
+
+#############################################################################
+# ------------
+#
+# References
+# """"""""""
+#
+# The use of the following functions, methods and classes is shown
+# in this example:
+
+matplotlib.dates.num2date
+matplotlib.dates.date2num
+matplotlib.dates.set_epoch
diff --git a/lib/matplotlib/axes/_axes.py b/lib/matplotlib/axes/_axes.py
@@ -1809,7 +1809,6 @@ def plot_date(self, x, y, fmt='o', tz=None, xdate=True, ydate=False,
             self.xaxis_date(tz)
         if ydate:
             self.yaxis_date(tz)
-
         ret = self.plot(x, y, fmt, **kwargs)
 
         self._request_autoscale_view()