def _windowize(x, w):
if not isinstance(x.op(), ops.WindowOp):
walked = _walk(x, w)
else:
window_arg, window_w = x.op().args
walked_child = _walk(window_arg, w)
if walked_child is not window_arg:
walked = x._factory(ops.WindowOp(walked_child, window_w),
name=x._name)
else:
walked = x
op = walked.op()
if isinstance(op, ops.AnalyticOp) or getattr(op, '_reduction', False):
if w is None:
w = window()
return walked.over(w)
elif isinstance(op, ops.WindowOp):
if w is not None:
return walked.over(w)
else:
return walked
else:
return walked
def _windowize(x, w):
if not isinstance(x.op(), ops.Window):
walked = _walk(x, w)
else:
window_arg, window_w = x.op().args
walked_child = _walk(window_arg, w)
if walked_child is not window_arg:
op = ops.Window(walked_child, window_w)
walked = op.to_expr().name(x.get_name())
else:
walked = x
op = walked.op()
if isinstance(op, (ops.Analytic, ops.Reduction)):
if w is None:
w = window()
return walked.over(w)
elif isinstance(op, ops.Window):
if w is not None:
return walked.over(w.combine(op.window))
else:
return walked
else:
return walked
def _windowize(x, w):
if not isinstance(x.op(), ops.WindowOp):
walked = _walk(x, w)
else:
window_arg, window_w = x.op().args
walked_child = _walk(window_arg, w)
if walked_child is not window_arg:
walked = x._factory(ops.WindowOp(walked_child, window_w),
name=x._name)
else:
walked = x
op = walked.op()
if (isinstance(op, ops.AnalyticOp) or
getattr(op, '_reduction', False)):
if w is None:
w = window()
return _check_window(walked.over(w))
elif isinstance(op, ops.WindowOp):
if w is not None:
return _check_window(walked.over(w))
else:
return _check_window(walked)
else:
return walked
def test_analytic_udf_destruct(backend, alltypes, udf):
w = window(preceding=None, following=None, group_by='year')
result = alltypes.mutate(
udf(alltypes['double_col'],
alltypes['int_col']).over(w).destructure()).execute()
expected = alltypes.mutate(
demean=alltypes['double_col'] - alltypes['double_col'].mean().over(w),
demean_weight=alltypes['int_col'] - alltypes['int_col'].mean().over(w),
).execute()
backend.assert_frame_equal(result, expected)
def _get_window(self):
if self._window is None:
groups = self.by
sorts = self._order_by
preceding, following = None, None
else:
w = self._window
groups = w.group_by + self.by
sorts = w.order_by + self._order_by
preceding, following = w.preceding, w.following
sorts = [ops.to_sort_key(self.table, k) for k in sorts]
return _window.window(preceding=preceding, following=following,
group_by=groups, order_by=sorts)
def _get_window(self):
if self._window is None:
groups = self.by
sorts = self._order_by
preceding, following = None, None
else:
w = self._window
groups = w.group_by + self.by
sorts = w.order_by + self._order_by
preceding, following = w.preceding, w.following
sorts = [ops.to_sort_key(self.table, k) for k in sorts]
return _window.window(preceding=preceding, following=following,
group_by=groups, order_by=sorts)
def test_analytic_udf_destruct_no_groupby(backend, alltypes):
w = window(preceding=None, following=None)
demean_struct_udf = create_demean_struct_udf(
result_formatter=lambda v1, v2: (v1, v2))
result = alltypes.mutate(
demean_struct_udf(
alltypes['double_col'],
alltypes['int_col']).over(w).destructure()).execute()
expected = alltypes.mutate(
demean=alltypes['double_col'] - alltypes['double_col'].mean().over(w),
demean_weight=alltypes['int_col'] - alltypes['int_col'].mean().over(w),
).execute()
backend.assert_frame_equal(result, expected)
def test_reduction_udf_destruct_window(backend, alltypes):
win = window(
preceding=ibis.interval(hours=2),
following=0,
group_by='year',
order_by='timestamp_col',
)
result = alltypes.mutate(
mean_struct(alltypes['double_col'],
alltypes['int_col']).over(win).destructure()).execute()
expected = alltypes.mutate(
mean=alltypes['double_col'].mean().over(win),
mean_weight=alltypes['int_col'].mean().over(win),
).execute()
backend.assert_frame_equal(result, expected)
def _get_window(self):
if self._window is None:
groups = self.by
sorts = self._order_by
preceding, following = None, None
else:
w = self._window
groups = w.group_by + self.by
sorts = w.order_by + self._order_by
preceding, following = w.preceding, w.following
return _window.window(
preceding=preceding,
following=following,
group_by=list(
map(self.table._ensure_expr, util.promote_list(groups))),
order_by=list(
map(self.table._ensure_expr, util.promote_list(sorts))),
)
def _get_window(self):
if self._window is None:
groups = self.by
sorts = self._order_by
preceding, following = None, None
else:
w = self._window
groups = w.group_by + self.by
sorts = w.order_by + self._order_by
preceding, following = w.preceding, w.following
sorts = [ops.sortkeys._to_sort_key(k, table=self.table) for k in sorts]
groups = _resolve_exprs(self.table, groups)
return _window.window(
preceding=preceding,
following=following,
group_by=groups,
order_by=sorts,
)
def test_analytic_udf_destruct_overwrite(backend, alltypes):
w = window(preceding=None, following=None, group_by='year')
result = alltypes.mutate(
overwrite_struct_analytic(
alltypes['double_col'],
alltypes['int_col']).over(w).destructure()).execute()
expected = alltypes.mutate(
double_col=alltypes['double_col'] -
alltypes['double_col'].mean().over(w),
demean_weight=alltypes['int_col'] - alltypes['int_col'].mean().over(w),
).execute()
# TODO issue #2649
# Due to a known limitation with how we treat DestructColumn
# in assignments, the ordering of op.selections may not exactly
# correspond with the column ordering we want (i.e. all new columns
# should appear at the end, but currently they are materialized
# directly after those overwritten columns).
backend.assert_frame_equal(result, expected, check_like=True)
def test_reduction_udf_destruct_window(udf_backend, udf_alltypes):
win = window(
preceding=ibis.interval(hours=2),
following=0,
group_by='year',
order_by='timestamp_col',
)
mean_struct_udf = create_mean_struct_udf(
result_formatter=lambda v1, v2: (v1, v2))
result = udf_alltypes.mutate(
mean_struct_udf(
udf_alltypes['double_col'],
udf_alltypes['int_col']).over(win).destructure()).execute()
expected = udf_alltypes.mutate(
mean=udf_alltypes['double_col'].mean().over(win),
mean_weight=udf_alltypes['int_col'].mean().over(win),
).execute()
udf_backend.assert_frame_equal(result, expected)
