ENH: use constant f32 eps, not np.finfo() during import

NumPy docs for np.finfo() say not to call it during import (at
module scope).  It's a relatively expensive call, and it modifies the
GIL state.  Now we just hard-code it, because it is always the value
anyway.  This avoids touching the GIL at import, which helps avoid
deadlocks in practice.

closes pandas-dev#14641

Author: John Zwinck <[email protected]>

Closes pandas-dev#15691 from jzwinck/patch-1 and squashes the following commits:

dadb97c [John Zwinck] DOC: mention pandas-dev#14641 in 0.20.0 whatsnew
e565230 [John Zwinck] ENH: use constant f32 eps, not np.finfo() during import

Author: jzwinck

doc/source/whatsnew/v0.20.0.txt

@@ -905,6 +905,7 @@ Bug Fixes
 
 
 - Bug in ``pd.read_csv()`` with ``float_precision='round_trip'`` which caused a segfault when a text entry is parsed (:issue:`15140`)
+- Avoid use of ``np.finfo()`` during ``import pandas`` removed to mitigate deadlock on Python GIL misuse (:issue:`14641`)
 
 - Bug in ``DataFrame.to_stata()`` and ``StataWriter`` which produces incorrectly formatted files to be produced for some locales (:issue:`13856`)
 - Bug in ``pd.concat()`` in which concatting with an empty dataframe with ``join='inner'`` was being improperly handled (:issue:`15328`)

pandas/core/indexing.py

@@ -1852,7 +1852,7 @@ def _convert_key(self, key, is_setter=False):
 
 
 # 32-bit floating point machine epsilon
-_eps = np.finfo('f4').eps
+_eps = 1.1920929e-07
 
 
 def length_of_indexer(indexer, target=None):