ENH: Fine-grained errstate handling

closes #13109
closes #13135

The precise strategy to be taken here is open for
discussion. I tried to be reasonably fine-grained rather than slap a
generic decorator over everything because it's easier to go that
direction than the reverse. The `errstate()` blocks in the tests were
added *after* fixing all of the library code. Unfortunately, these are
less fine-grained than I would like because some of the tests have
many lines of the form `assert_array_equal(pandas_expression_to_test,
expected_raw_numpy_expression)` where `expected_raw_numpy_expression`
is what is triggering the warning. It was tedious to try to rewrite
all of that to wrap just `expected_raw_numpy_expression`.    I think I
got everything exercised by the test suite except for parts of the
test suite that are skipped on my machine due to dependencies. We'll
see how things go in the CI.    I haven't added any new tests yet.
Could do if requested.

Author: Robert Kern <[email protected]>
Author: Robert Kern <[email protected]>

Closes #13145 from rkern/fix/errstate and squashes the following commits:

ef9c001 [Robert Kern] BUG: whoops, wrong function.
7fd2e86 [Robert Kern] ENH: More whatsnew documentation.
44805db [Robert Kern] ENH: Rearrange expression to avoid generating a warning that would need to be silenced.
1fe1bc2 [Robert Kern] pep8
bf1f662 [Robert Kern] BUG: New fixes after master rebase.
e7adc03 [Robert Kern] BUG: wrong function.
a59cfa7 [Robert Kern] ENH: Avoiding the bounds error is better than silencing the warning.
0e1ea81 [Robert Kern] BUG: A few more stragglers.
863ac93 [Robert Kern] TST: Add a new test to ensure that boolean comparisons are errstate-protected.
6932851 [Robert Kern] TST: Basic check that the global errstate remains unchanged.
c9df7b3 [Robert Kern] BUG: removed debugging print
3b12f08 [Robert Kern] ENH: Silence numpy warnings from certain expressions computed during tests.
eca512c [Robert Kern] BUG: Handle NaT explicitly.
6fbc9ce [Robert Kern] BUG: First pass at fine-grained errstate.

Author: rkern

doc/source/whatsnew/v0.19.0.txt

@@ -7,6 +7,10 @@ This is a major release from 0.18.1 and includes a small number of API changes,
 enhancements, and performance improvements along with a large number of bug fixes. We recommend that all
 users upgrade to this version.
 
+.. warning::
+
+    pandas >= 0.19.0 will no longer silence numpy ufunc warnings upon import, see :ref:`here <whatsnew_0190.errstate>`. (:issue:`13109`, :issue:`13145`)
+
 Highlights include:
 
 - :func:`merge_asof` for asof-style time-series joining, see :ref:`here <whatsnew_0190.enhancements.asof_merge>`
@@ -357,6 +361,15 @@ Google BigQuery Enhancements
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 - The :func:`pandas.io.gbq.read_gbq` method has gained the ``dialect`` argument to allow users to specify whether to use BigQuery's legacy SQL or BigQuery's standard SQL. See the :ref:`docs <io.bigquery_reader>` for more details (:issue:`13615`).
 
+.. _whatsnew_0190.errstate:
+
+Fine-grained numpy errstate
+^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Previous versions of pandas would permanently silence numpy's ufunc error handling when ``pandas`` was imported (:issue:`13109`). Pandas did this in order to silence the warnings that would arise from using numpy ufuncs on missing data, which are usually represented as NaNs. Unfortunately, this silenced legitimate warnings arising in non-pandas code in the application. Starting with 0.19.0, pandas will use the ``numpy.errstate`` context manager to silence these warnings in a more fine-grained manner only around where these operations are actually used in the pandas codebase.
+
+After upgrading pandas, you may see "new" ``RuntimeWarnings`` being issued from your code. These are likely legitimate, and the underlying cause likely existed in the code when using previous versions of pandas that simply silenced the warning. Use `numpy.errstate <http://docs.scipy.org/doc/numpy/reference/generated/numpy.errstate.html>`__ around the source of the ``RuntimeWarning`` to control how these conditions are handled.
+
 .. _whatsnew_0190.enhancements.other:
 
 Other enhancements

pandas/compat/numpy/__init__.py

@@ -5,8 +5,6 @@
 from distutils.version import LooseVersion
 from pandas.compat import string_types, string_and_binary_types
 
-# turn off all numpy warnings
-np.seterr(all='ignore')
 
 # numpy versioning
 _np_version = np.version.short_version

pandas/computation/align.py

@@ -95,7 +95,7 @@ def _align_core(terms):
                 term_axis_size = len(ti.axes[axis])
                 reindexer_size = len(reindexer)
 
-                ordm = np.log10(abs(reindexer_size - term_axis_size))
+                ordm = np.log10(max(1, abs(reindexer_size - term_axis_size)))
                 if ordm >= 1 and reindexer_size >= 10000:
                     warnings.warn('Alignment difference on axis {0} is larger '
                                   'than an order of magnitude on term {1!r}, '

pandas/computation/expressions.py

@@ -59,7 +59,8 @@ def _evaluate_standard(op, op_str, a, b, raise_on_error=True, **eval_kwargs):
     """ standard evaluation """
     if _TEST_MODE:
         _store_test_result(False)
-    return op(a, b)
+    with np.errstate(all='ignore'):
+        return op(a, b)
 
 
 def _can_use_numexpr(op, op_str, a, b, dtype_check):

pandas/computation/ops.py

@@ -523,7 +523,8 @@ def __init__(self, func, args):
 
     def __call__(self, env):
         operands = [op(env) for op in self.operands]
-        return self.func.func(*operands)
+        with np.errstate(all='ignore'):
+            return self.func.func(*operands)
 
     def __unicode__(self):
         operands = map(str, self.operands)

pandas/computation/tests/test_eval.py

@@ -1613,7 +1613,8 @@ def test_unary_functions(self):
         for fn in self.unary_fns:
             expr = "{0}(a)".format(fn)
             got = self.eval(expr)
-            expect = getattr(np, fn)(a)
+            with np.errstate(all='ignore'):
+                expect = getattr(np, fn)(a)
             tm.assert_series_equal(got, expect, check_names=False)
 
     def test_binary_functions(self):
@@ -1624,7 +1625,8 @@ def test_binary_functions(self):
         for fn in self.binary_fns:
             expr = "{0}(a, b)".format(fn)
             got = self.eval(expr)
-            expect = getattr(np, fn)(a, b)
+            with np.errstate(all='ignore'):
+                expect = getattr(np, fn)(a, b)
             tm.assert_almost_equal(got, expect, check_names=False)
 
     def test_df_use_case(self):

pandas/core/frame.py

@@ -3810,7 +3810,8 @@ def update(self, other, join='left', overwrite=True, filter_func=None,
             this = self[col].values
             that = other[col].values
             if filter_func is not None:
-                mask = ~filter_func(this) | isnull(that)
+                with np.errstate(all='ignore'):
+                    mask = ~filter_func(this) | isnull(that)
             else:
                 if raise_conflict:
                     mask_this = notnull(that)
@@ -4105,7 +4106,8 @@ def f(x):
             return self._apply_empty_result(func, axis, reduce, *args, **kwds)
 
         if isinstance(f, np.ufunc):
-            results = f(self.values)
+            with np.errstate(all='ignore'):
+                results = f(self.values)
             return self._constructor(data=results, index=self.index,
                                      columns=self.columns, copy=False)
         else:
@@ -4931,7 +4933,8 @@ def f(x):
                                             "type %s not implemented." %
                                             filter_type)
                     raise_with_traceback(e)
-                result = f(data.values)
+                with np.errstate(all='ignore'):
+                    result = f(data.values)
                 labels = data._get_agg_axis(axis)
         else:
             if numeric_only:

pandas/core/groupby.py

@@ -678,7 +678,8 @@ def apply(self, func, *args, **kwargs):
 
                 @wraps(func)
                 def f(g):
-                    return func(g, *args, **kwargs)
+                    with np.errstate(all='ignore'):
+                        return func(g, *args, **kwargs)
             else:
                 raise ValueError('func must be a callable if args or '
                                  'kwargs are supplied')
@@ -4126,7 +4127,10 @@ def loop(labels, shape):
         out = stride * labels[0].astype('i8', subok=False, copy=False)
 
         for i in range(1, nlev):
-            stride //= shape[i]
+            if shape[i] == 0:
+                stride = 0
+            else:
+                stride //= shape[i]
             out += labels[i] * stride
 
         if xnull:  # exclude nulls
@@ -4365,7 +4369,9 @@ def _get_group_index_sorter(group_index, ngroups):
     count = len(group_index)
     alpha = 0.0  # taking complexities literally; there may be
     beta = 1.0  # some room for fine-tuning these parameters
-    if alpha + beta * ngroups < count * np.log(count):
+    do_groupsort = (count > 0 and ((alpha + beta * ngroups) <
+                                   (count * np.log(count))))
+    if do_groupsort:
         sorter, _ = _algos.groupsort_indexer(_ensure_int64(group_index),
                                              ngroups)
         return _ensure_platform_int(sorter)

pandas/core/internals.py

@@ -348,7 +348,8 @@ def apply(self, func, mgr=None, **kwargs):
         """ apply the function to my values; return a block if we are not
         one
         """
-        result = func(self.values, **kwargs)
+        with np.errstate(all='ignore'):
+            result = func(self.values, **kwargs)
         if not isinstance(result, Block):
             result = self.make_block(values=_block_shape(result,
                                                          ndim=self.ndim))
@@ -1156,7 +1157,8 @@ def handle_error():
 
         # get the result
         try:
-            result = get_result(other)
+            with np.errstate(all='ignore'):
+                result = get_result(other)
 
         # if we have an invalid shape/broadcast error
         # GH4576, so raise instead of allowing to pass through

pandas/core/nanops.py

@@ -45,7 +45,8 @@ def _f(*args, **kwargs):
                                 'this dtype'.format(
                                     f.__name__.replace('nan', '')))
             try:
-                return f(*args, **kwargs)
+                with np.errstate(invalid='ignore'):
+                    return f(*args, **kwargs)
             except ValueError as e:
                 # we want to transform an object array
                 # ValueError message to the more typical TypeError
@@ -513,7 +514,8 @@ def nanskew(values, axis=None, skipna=True):
     m2 = _zero_out_fperr(m2)
     m3 = _zero_out_fperr(m3)
 
-    result = (count * (count - 1) ** 0.5 / (count - 2)) * (m3 / m2 ** 1.5)
+    with np.errstate(invalid='ignore', divide='ignore'):
+        result = (count * (count - 1) ** 0.5 / (count - 2)) * (m3 / m2 ** 1.5)
 
     dtype = values.dtype
     if is_float_dtype(dtype):
@@ -562,10 +564,11 @@ def nankurt(values, axis=None, skipna=True):
     m2 = adjusted2.sum(axis, dtype=np.float64)
     m4 = adjusted4.sum(axis, dtype=np.float64)
 
-    adj = 3 * (count - 1) ** 2 / ((count - 2) * (count - 3))
-    numer = count * (count + 1) * (count - 1) * m4
-    denom = (count - 2) * (count - 3) * m2**2
-    result = numer / denom - adj
+    with np.errstate(invalid='ignore', divide='ignore'):
+        adj = 3 * (count - 1) ** 2 / ((count - 2) * (count - 3))
+        numer = count * (count + 1) * (count - 1) * m4
+        denom = (count - 2) * (count - 3) * m2**2
+        result = numer / denom - adj
 
     # floating point error
     numer = _zero_out_fperr(numer)
@@ -579,7 +582,8 @@ def nankurt(values, axis=None, skipna=True):
         if denom == 0:
             return 0
 
-    result = numer / denom - adj
+    with np.errstate(invalid='ignore', divide='ignore'):
+        result = numer / denom - adj
 
     dtype = values.dtype
     if is_float_dtype(dtype):
@@ -658,7 +662,8 @@ def _maybe_null_out(result, axis, mask):
 
 def _zero_out_fperr(arg):
     if isinstance(arg, np.ndarray):
-        return np.where(np.abs(arg) < 1e-14, 0, arg)
+        with np.errstate(invalid='ignore'):
+            return np.where(np.abs(arg) < 1e-14, 0, arg)
     else:
         return arg.dtype.type(0) if np.abs(arg) < 1e-14 else arg
 
@@ -760,7 +765,8 @@ def f(x, y):
         ymask = isnull(y)
         mask = xmask | ymask
 
-        result = op(x, y)
+        with np.errstate(all='ignore'):
+            result = op(x, y)
 
         if mask.any():
             if is_bool_dtype(result):