FIX: revert part of pandas-dev#11770

Reverts part of:

pandas-dev#11770

Closes:

matplotlib/matplotlib#6365

The mistake in pandas-dev#11770 was missing that pandas had a 1/us not 1/s
scaled bucket.

Author: tacaswell

doc/source/whatsnew/v0.18.2.txt

@@ -143,7 +143,4 @@ Bug Fixes
 
 
 
-
-
-
-
+- Bug in ``AutoDataFormatter``, restore second scaled formatting and readd musecond scaled formatting (:issue:`13131`)

pandas/tseries/converter.py

@@ -4,6 +4,7 @@
 
 from dateutil.relativedelta import relativedelta
 
+import matplotlib
 import matplotlib.units as units
 import matplotlib.dates as dates
 
@@ -23,6 +24,16 @@
 from pandas.tseries.frequencies import FreqGroup
 from pandas.tseries.period import Period, PeriodIndex
 
+# constants
+HOURS_PER_DAY = 24.
+MIN_PER_HOUR = 60.
+SEC_PER_MIN = 60.
+
+SEC_PER_HOUR = SEC_PER_MIN * MIN_PER_HOUR
+SEC_PER_DAY = SEC_PER_HOUR * HOURS_PER_DAY
+
+MUSEC_PER_DAY = 1e6 * SEC_PER_DAY
+
 
 def register():
     units.registry[lib.Timestamp] = DatetimeConverter()
@@ -221,6 +232,13 @@ def __init__(self, locator, tz=None, defaultfmt='%Y-%m-%d'):
         if self._tz is dates.UTC:
             self._tz._utcoffset = self._tz.utcoffset(None)
 
+        # For mpl > 2.0 the format strings are controlled via rcparams
+        # so do not mess with them.  For mpl < 2.0 change the second
+        # break point and add a musec break point
+        if matplotlib.compare_versions('2.0.0', matplotlib.__version__):
+            self.scaled[1. / SEC_PER_DAY] = '%H:%M:%S'
+            self.scaled[1. / MUSEC_PER_DAY] = '%H:%M:%S.%f'
+
 
 class PandasAutoDateLocator(dates.AutoDateLocator):