def exec_post_processing(self, df: DataFrame) -> DataFrame:
"""
Perform post processing operations on DataFrame.
:param df: DataFrame returned from database model.
:return: new DataFrame to which all post processing operations have been
applied
:raises QueryObjectValidationError: If the post processing operation
is incorrect
"""
logger.debug("post_processing: \n %s", pformat(self.post_processing))
for post_process in self.post_processing:
operation = post_process.get("operation")
if not operation:
raise InvalidPostProcessingError(
_("`operation` property of post processing object undefined")
)
if not hasattr(pandas_postprocessing, operation):
raise InvalidPostProcessingError(
_(
"Unsupported post processing operation: %(operation)s",
type=operation,
)
)
options = post_process.get("options", {})
df = getattr(pandas_postprocessing, operation)(df, **options)
return df
def resample(
df: pd.DataFrame,
rule: str,
method: str,
fill_value: Optional[Union[float, int]] = None,
) -> pd.DataFrame:
"""
support upsampling in resample
:param df: DataFrame to resample.
:param rule: The offset string representing target conversion.
:param method: How to fill the NaN value after resample.
:param fill_value: What values do fill missing.
:return: DataFrame after resample
:raises InvalidPostProcessingError: If the request in incorrect
"""
if not isinstance(df.index, pd.DatetimeIndex):
raise InvalidPostProcessingError(_("Resample operation requires DatetimeIndex"))
if method not in RESAMPLE_METHOD:
raise InvalidPostProcessingError(
_("Resample method should in ") + ", ".join(RESAMPLE_METHOD) + "."
)
if method == "asfreq" and fill_value is not None:
_df = df.resample(rule).asfreq(fill_value=fill_value)
elif method == "linear":
_df = df.resample(rule).interpolate()
else:
_df = getattr(df.resample(rule), method)()
return _df
def compare( # pylint: disable=too-many-arguments
df: DataFrame,
source_columns: List[str],
compare_columns: List[str],
compare_type: PandasPostprocessingCompare,
drop_original_columns: Optional[bool] = False,
precision: Optional[int] = 4,
) -> DataFrame:
"""
Calculate column-by-column changing for select columns.
:param df: DataFrame on which the compare will be based.
:param source_columns: Main query columns
:param compare_columns: Columns being compared
:param compare_type: Type of compare. Choice of `absolute`, `percentage` or `ratio`
:param drop_original_columns: Whether to remove the source columns and
compare columns.
:param precision: Round a change rate to a variable number of decimal places.
:return: DataFrame with compared columns.
:raises InvalidPostProcessingError: If the request in incorrect.
"""
if len(source_columns) != len(compare_columns):
raise InvalidPostProcessingError(
_("`compare_columns` must have the same length as `source_columns`."
))
if compare_type not in tuple(PandasPostprocessingCompare):
raise InvalidPostProcessingError(
_("`compare_type` must be `difference`, `percentage` or `ratio`"))
if len(source_columns) == 0:
return df
for s_col, c_col in zip(source_columns, compare_columns):
s_df = df.loc[:, [s_col]]
s_df.rename(columns={s_col: "__intermediate"}, inplace=True)
c_df = df.loc[:, [c_col]]
c_df.rename(columns={c_col: "__intermediate"}, inplace=True)
if compare_type == PandasPostprocessingCompare.DIFF:
diff_df = s_df - c_df
elif compare_type == PandasPostprocessingCompare.PCT:
diff_df = ((s_df - c_df) / c_df).astype(float).round(precision)
else:
# compare_type == "ratio"
diff_df = (s_df / c_df).astype(float).round(precision)
diff_df.rename(
columns={
"__intermediate":
TIME_COMPARISON.join([compare_type, s_col, c_col])
},
inplace=True,
)
df = pd.concat([df, diff_df], axis=1)
if drop_original_columns:
df = df.drop(source_columns + compare_columns, axis=1)
return df
def contribution(
df: DataFrame,
orientation: Optional[
PostProcessingContributionOrientation
] = PostProcessingContributionOrientation.COLUMN,
columns: Optional[List[str]] = None,
rename_columns: Optional[List[str]] = None,
) -> DataFrame:
"""
Calculate cell contibution to row/column total for numeric columns.
Non-numeric columns will be kept untouched.
If `columns` are specified, only calculate contributions on selected columns.
:param df: DataFrame containing all-numeric data (temporal column ignored)
:param columns: Columns to calculate values from.
:param rename_columns: The new labels for the calculated contribution columns.
The original columns will not be removed.
:param orientation: calculate by dividing cell with row/column total
:return: DataFrame with contributions.
"""
contribution_df = df.copy()
numeric_df = contribution_df.select_dtypes(include=["number", Decimal])
# TODO: copy needed due to following regression in 1.4, remove if not needed:
# https://github.com/pandas-dev/pandas/issues/48090
numeric_df = numeric_df.copy()
numeric_df.fillna(0, inplace=True)
# verify column selections
if columns:
numeric_columns = numeric_df.columns.tolist()
for col in columns:
if col not in numeric_columns:
raise InvalidPostProcessingError(
_(
'Column "%(column)s" is not numeric or does not '
"exists in the query results.",
column=col,
)
)
columns = columns or numeric_df.columns
rename_columns = rename_columns or columns
if len(rename_columns) != len(columns):
raise InvalidPostProcessingError(
_("`rename_columns` must have the same length as `columns`.")
)
# limit to selected columns
numeric_df = numeric_df[columns]
axis = 0 if orientation == PostProcessingContributionOrientation.COLUMN else 1
numeric_df = numeric_df / numeric_df.values.sum(axis=axis, keepdims=True)
contribution_df[rename_columns] = numeric_df
return contribution_df
def cum(
df: DataFrame,
operator: str,
columns: Dict[str, str],
) -> DataFrame:
"""
Calculate cumulative sum/product/min/max for select columns.
:param df: DataFrame on which the cumulative operation will be based.
:param columns: columns on which to perform a cumulative operation, mapping source
column to target column. For instance, `{'y': 'y'}` will replace the column
`y` with the cumulative value in `y`, while `{'y': 'y2'}` will add a column
`y2` based on cumulative values calculated from `y`, leaving the original
column `y` unchanged.
:param operator: cumulative operator, e.g. `sum`, `prod`, `min`, `max`
:return: DataFrame with cumulated columns
"""
columns = columns or {}
df_cum = df.loc[:, columns.keys()]
operation = "cum" + operator
if operation not in ALLOWLIST_CUMULATIVE_FUNCTIONS or not hasattr(
df_cum, operation):
raise InvalidPostProcessingError(
_("Invalid cumulative operator: %(operator)s", operator=operator))
df_cum = _append_columns(df, getattr(df_cum, operation)(), columns)
return df_cum
def geohash_encode(
df: DataFrame,
geohash: str,
longitude: str,
latitude: str,
) -> DataFrame:
"""
Encode longitude and latitude into geohash
:param df: DataFrame containing longitude and latitude data
:param geohash: Name of new column to be created containing geohash location.
:param longitude: Name of source column containing longitude.
:param latitude: Name of source column containing latitude.
:return: DataFrame with decoded longitudes and latitudes
"""
try:
encode_df = df[[latitude, longitude]]
encode_df.columns = ["latitude", "longitude"]
encode_df["geohash"] = encode_df.apply(
lambda row: geohash_lib.encode(row["latitude"], row["longitude"]),
axis=1,
)
return _append_columns(df, encode_df, {"geohash": geohash})
except ValueError as ex:
raise InvalidPostProcessingError(
_("Invalid longitude/latitude")) from ex
def _get_aggregate_funcs(
df: DataFrame,
aggregates: Dict[str, Dict[str, Any]],
) -> Dict[str, NamedAgg]:
"""
Converts a set of aggregate config objects into functions that pandas can use as
aggregators. Currently only numpy aggregators are supported.
:param df: DataFrame on which to perform aggregate operation.
:param aggregates: Mapping from column name to aggregate config.
:return: Mapping from metric name to function that takes a single input argument.
"""
agg_funcs: Dict[str, NamedAgg] = {}
for name, agg_obj in aggregates.items():
column = agg_obj.get("column", name)
if column not in df:
raise InvalidPostProcessingError(
_(
"Column referenced by aggregate is undefined: %(column)s",
column=column,
)
)
if "operator" not in agg_obj:
raise InvalidPostProcessingError(
_(
"Operator undefined for aggregator: %(name)s",
name=name,
)
)
operator = agg_obj["operator"]
if callable(operator):
aggfunc = operator
else:
func = NUMPY_FUNCTIONS.get(operator)
if not func:
raise InvalidPostProcessingError(
_(
"Invalid numpy function: %(operator)s",
operator=operator,
)
)
options = agg_obj.get("options", {})
aggfunc = partial(func, **options)
agg_funcs[name] = NamedAgg(column=column, aggfunc=aggfunc)
return agg_funcs
def wrapped(df: DataFrame, **options: Any) -> Any:
if _is_multi_index_on_columns(df):
# MultiIndex column validate first level
columns = df.columns.get_level_values(0)
else:
columns = df.columns.tolist()
for name in argnames:
if name in options and not all(
elem in columns for elem in options.get(name) or []):
raise InvalidPostProcessingError(
_("Referenced columns not available in DataFrame."))
return func(df, **options)
def geohash_decode(df: DataFrame, geohash: str, longitude: str,
latitude: str) -> DataFrame:
"""
Decode a geohash column into longitude and latitude
:param df: DataFrame containing geohash data
:param geohash: Name of source column containing geohash location.
:param longitude: Name of new column to be created containing longitude.
:param latitude: Name of new column to be created containing latitude.
:return: DataFrame with decoded longitudes and latitudes
"""
try:
lonlat_df = DataFrame()
lonlat_df["latitude"], lonlat_df["longitude"] = zip(
*df[geohash].apply(geohash_lib.decode))
return _append_columns(df, lonlat_df, {
"latitude": latitude,
"longitude": longitude
})
except ValueError as ex:
raise InvalidPostProcessingError(_("Invalid geohash string")) from ex
def geodetic_parse(
df: DataFrame,
geodetic: str,
longitude: str,
latitude: str,
altitude: Optional[str] = None,
) -> DataFrame:
"""
Parse a column containing a geodetic point string
[Geopy](https://geopy.readthedocs.io/en/stable/#geopy.point.Point).
:param df: DataFrame containing geodetic point data
:param geodetic: Name of source column containing geodetic point string.
:param longitude: Name of new column to be created containing longitude.
:param latitude: Name of new column to be created containing latitude.
:param altitude: Name of new column to be created containing altitude.
:return: DataFrame with decoded longitudes and latitudes
"""
def _parse_location(location: str) -> Tuple[float, float, float]:
"""
Parse a string containing a geodetic point and return latitude, longitude
and altitude
"""
point = Point(location)
return point[0], point[1], point[2]
try:
geodetic_df = DataFrame()
(
geodetic_df["latitude"],
geodetic_df["longitude"],
geodetic_df["altitude"],
) = zip(*df[geodetic].apply(_parse_location))
columns = {"latitude": latitude, "longitude": longitude}
if altitude:
columns["altitude"] = altitude
return _append_columns(df, geodetic_df, columns)
except ValueError as ex:
raise InvalidPostProcessingError(_("Invalid geodetic string")) from ex
def _prophet_fit_and_predict( # pylint: disable=too-many-arguments
df: DataFrame,
confidence_interval: float,
yearly_seasonality: Union[bool, str, int],
weekly_seasonality: Union[bool, str, int],
daily_seasonality: Union[bool, str, int],
periods: int,
freq: str,
) -> DataFrame:
"""
Fit a prophet model and return a DataFrame with predicted results.
"""
try:
# pylint: disable=import-error,import-outside-toplevel
from prophet import Prophet
prophet_logger = logging.getLogger("prophet.plot")
prophet_logger.setLevel(logging.CRITICAL)
prophet_logger.setLevel(logging.NOTSET)
except ModuleNotFoundError as ex:
raise InvalidPostProcessingError(
_("`prophet` package not installed")) from ex
model = Prophet(
interval_width=confidence_interval,
yearly_seasonality=yearly_seasonality,
weekly_seasonality=weekly_seasonality,
daily_seasonality=daily_seasonality,
)
if df["ds"].dt.tz:
df["ds"] = df["ds"].dt.tz_convert(None)
model.fit(df)
future = model.make_future_dataframe(periods=periods, freq=freq)
forecast = model.predict(future)[[
"ds", "yhat", "yhat_lower", "yhat_upper"
]]
return forecast.join(df.set_index("ds"), on="ds").set_index(["ds"])
def pivot( # pylint: disable=too-many-arguments,too-many-locals
df: DataFrame,
index: List[str],
aggregates: Dict[str, Dict[str, Any]],
columns: Optional[List[str]] = None,
metric_fill_value: Optional[Any] = None,
column_fill_value: Optional[str] = NULL_STRING,
drop_missing_columns: Optional[bool] = True,
combine_value_with_metric: bool = False,
marginal_distributions: Optional[bool] = None,
marginal_distribution_name: Optional[str] = None,
flatten_columns: bool = True,
reset_index: bool = True,
) -> DataFrame:
"""
Perform a pivot operation on a DataFrame.
:param df: Object on which pivot operation will be performed
:param index: Columns to group by on the table index (=rows)
:param columns: Columns to group by on the table columns
:param metric_fill_value: Value to replace missing values with
:param column_fill_value: Value to replace missing pivot columns with. By default
replaces missing values with "<NULL>". Set to `None` to remove columns
with missing values.
:param drop_missing_columns: Do not include columns whose entries are all missing
:param combine_value_with_metric: Display metrics side by side within each column,
as opposed to each column being displayed side by side for each metric.
:param aggregates: A mapping from aggregate column name to the aggregate
config.
:param marginal_distributions: Add totals for row/column. Default to False
:param marginal_distribution_name: Name of row/column with marginal distribution.
Default to 'All'.
:param flatten_columns: Convert column names to strings
:param reset_index: Convert index to column
:return: A pivot table
:raises InvalidPostProcessingError: If the request in incorrect
"""
if not index:
raise InvalidPostProcessingError(
_("Pivot operation requires at least one index"))
if not aggregates:
raise InvalidPostProcessingError(
_("Pivot operation must include at least one aggregate"))
if columns and column_fill_value:
df[columns] = df[columns].fillna(value=column_fill_value)
aggregate_funcs = _get_aggregate_funcs(df, aggregates)
# TODO (villebro): Pandas 1.0.3 doesn't yet support NamedAgg in pivot_table.
# Remove once/if support is added.
aggfunc = {na.column: na.aggfunc for na in aggregate_funcs.values()}
# When dropna = False, the pivot_table function will calculate cartesian-product
# for MultiIndex.
# https://github.com/apache/superset/issues/15956
# https://github.com/pandas-dev/pandas/issues/18030
series_set = set()
if not drop_missing_columns and columns:
for row in df[columns].itertuples():
for metric in aggfunc.keys():
series_set.add(str(tuple([metric]) + tuple(row[1:])))
df = df.pivot_table(
values=aggfunc.keys(),
index=index,
columns=columns,
aggfunc=aggfunc,
fill_value=metric_fill_value,
dropna=drop_missing_columns,
margins=marginal_distributions,
margins_name=marginal_distribution_name,
)
if not drop_missing_columns and len(series_set) > 0 and not df.empty:
for col in df.columns:
series = str(col)
if series not in series_set:
df = df.drop(col, axis=PandasAxis.COLUMN)
if combine_value_with_metric:
df = df.stack(0).unstack()
# Make index regular column
if flatten_columns:
df.columns = [
_flatten_column_after_pivot(col, aggregates) for col in df.columns
]
# return index as regular column
if reset_index:
df.reset_index(level=0, inplace=True)
return df
def boxplot(
df: DataFrame,
groupby: List[str],
metrics: List[str],
whisker_type: PostProcessingBoxplotWhiskerType,
percentiles: Optional[Union[List[Union[int, float]],
Tuple[Union[int, float],
Union[int, float]]]] = None,
) -> DataFrame:
"""
Calculate boxplot statistics. For each metric, the operation creates eight
new columns with the column name suffixed with the following values:
- `__mean`: the mean
- `__median`: the median
- `__max`: the maximum value excluding outliers (see whisker type)
- `__min`: the minimum value excluding outliers (see whisker type)
- `__q1`: the median
- `__q1`: the first quartile (25th percentile)
- `__q3`: the third quartile (75th percentile)
- `__count`: count of observations
- `__outliers`: the values that fall outside the minimum/maximum value
(see whisker type)
:param df: DataFrame containing all-numeric data (temporal column ignored)
:param groupby: The categories to group by (x-axis)
:param metrics: The metrics for which to calculate the distribution
:param whisker_type: The confidence level type
:return: DataFrame with boxplot statistics per groupby
"""
def quartile1(series: Series) -> float:
return np.nanpercentile(series, 25, interpolation="midpoint")
def quartile3(series: Series) -> float:
return np.nanpercentile(series, 75, interpolation="midpoint")
if whisker_type == PostProcessingBoxplotWhiskerType.TUKEY:
def whisker_high(series: Series) -> float:
upper_outer_lim = quartile3(series) + 1.5 * (quartile3(series) -
quartile1(series))
return series[series <= upper_outer_lim].max()
def whisker_low(series: Series) -> float:
lower_outer_lim = quartile1(series) - 1.5 * (quartile3(series) -
quartile1(series))
return series[series >= lower_outer_lim].min()
elif whisker_type == PostProcessingBoxplotWhiskerType.PERCENTILE:
if (not isinstance(percentiles, (list, tuple)) or len(percentiles) != 2
or not isinstance(percentiles[0], (int, float))
or not isinstance(percentiles[1], (int, float))
or percentiles[0] >= percentiles[1]):
raise InvalidPostProcessingError(
_("percentiles must be a list or tuple with two numeric values, "
"of which the first is lower than the second value"))
low, high = percentiles[0], percentiles[1]
def whisker_high(series: Series) -> float:
return np.nanpercentile(series, high)
def whisker_low(series: Series) -> float:
return np.nanpercentile(series, low)
else:
whisker_high = np.max
whisker_low = np.min
def outliers(series: Series) -> Set[float]:
above = series[series > whisker_high(series)]
below = series[series < whisker_low(series)]
return above.tolist() + below.tolist()
operators: Dict[str, Callable[[Any], Any]] = {
"mean": np.mean,
"median": np.median,
"max": whisker_high,
"min": whisker_low,
"q1": quartile1,
"q3": quartile3,
"count": np.ma.count,
"outliers": outliers,
}
aggregates: Dict[str, Dict[str, Union[str, Callable[..., Any]]]] = {
f"{metric}__{operator_name}": {
"column": metric,
"operator": operator
}
for operator_name, operator in operators.items() for metric in metrics
}
# nanpercentile needs numeric values, otherwise the isnan function
# that's used in the underlying function will fail
for column in metrics:
if df.dtypes[column] == np.object:
df[column] = to_numeric(df[column], errors="coerce")
return aggregate(df, groupby=groupby, aggregates=aggregates)
def prophet( # pylint: disable=too-many-arguments
df: DataFrame,
time_grain: str,
periods: int,
confidence_interval: float,
yearly_seasonality: Optional[Union[bool, int]] = None,
weekly_seasonality: Optional[Union[bool, int]] = None,
daily_seasonality: Optional[Union[bool, int]] = None,
index: Optional[str] = None,
) -> DataFrame:
"""
Add forecasts to each series in a timeseries dataframe, along with confidence
intervals for the prediction. For each series, the operation creates three
new columns with the column name suffixed with the following values:
- `__yhat`: the forecast for the given date
- `__yhat_lower`: the lower bound of the forecast for the given date
- `__yhat_upper`: the upper bound of the forecast for the given date
:param df: DataFrame containing all-numeric data (temporal column ignored)
:param time_grain: Time grain used to specify time period increments in prediction
:param periods: Time periods (in units of `time_grain`) to predict into the future
:param confidence_interval: Width of predicted confidence interval
:param yearly_seasonality: Should yearly seasonality be applied.
An integer value will specify Fourier order of seasonality.
:param weekly_seasonality: Should weekly seasonality be applied.
An integer value will specify Fourier order of seasonality, `None` will
automatically detect seasonality.
:param daily_seasonality: Should daily seasonality be applied.
An integer value will specify Fourier order of seasonality, `None` will
automatically detect seasonality.
:param index: the name of the column containing the x-axis data
:return: DataFrame with contributions, with temporal column at beginning if present
"""
index = index or DTTM_ALIAS
# validate inputs
if not time_grain:
raise InvalidPostProcessingError(_("Time grain missing"))
if time_grain not in PROPHET_TIME_GRAIN_MAP:
raise InvalidPostProcessingError(
_(
"Unsupported time grain: %(time_grain)s",
time_grain=time_grain,
))
freq = PROPHET_TIME_GRAIN_MAP[time_grain]
# check type at runtime due to marhsmallow schema not being able to handle
# union types
if not isinstance(periods, int) or periods < 0:
raise InvalidPostProcessingError(_("Periods must be a whole number"))
if not confidence_interval or confidence_interval <= 0 or confidence_interval >= 1:
raise InvalidPostProcessingError(
_("Confidence interval must be between 0 and 1 (exclusive)"))
if index not in df.columns:
raise InvalidPostProcessingError(
_("DataFrame must include temporal column"))
if len(df.columns) < 2:
raise InvalidPostProcessingError(
_("DataFrame include at least one series"))
target_df = DataFrame()
for column in [column for column in df.columns if column != index]:
fit_df = _prophet_fit_and_predict(
df=df[[index, column]].rename(columns={
index: "ds",
column: "y"
}),
confidence_interval=confidence_interval,
yearly_seasonality=_prophet_parse_seasonality(yearly_seasonality),
weekly_seasonality=_prophet_parse_seasonality(weekly_seasonality),
daily_seasonality=_prophet_parse_seasonality(daily_seasonality),
periods=periods,
freq=freq,
)
new_columns = [
f"{column}__yhat",
f"{column}__yhat_lower",
f"{column}__yhat_upper",
f"{column}",
]
fit_df.columns = new_columns
if target_df.empty:
target_df = fit_df
else:
for new_column in new_columns:
target_df = target_df.assign(
**{new_column: fit_df[new_column]})
target_df.reset_index(level=0, inplace=True)
return target_df.rename(columns={"ds": index})
def rolling( # pylint: disable=too-many-arguments
df: DataFrame,
rolling_type: str,
columns: Dict[str, str],
window: Optional[int] = None,
rolling_type_options: Optional[Dict[str, Any]] = None,
center: bool = False,
win_type: Optional[str] = None,
min_periods: Optional[int] = None,
) -> DataFrame:
"""
Apply a rolling window on the dataset. See the Pandas docs for further details:
https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.rolling.html
:param df: DataFrame on which the rolling period will be based.
:param columns: columns on which to perform rolling, mapping source column to
target column. For instance, `{'y': 'y'}` will replace the column `y` with
the rolling value in `y`, while `{'y': 'y2'}` will add a column `y2` based
on rolling values calculated from `y`, leaving the original column `y`
unchanged.
:param rolling_type: Type of rolling window. Any numpy function will work.
:param window: Size of the window.
:param rolling_type_options: Optional options to pass to rolling method. Needed
for e.g. quantile operation.
:param center: Should the label be at the center of the window.
:param win_type: Type of window function.
:param min_periods: The minimum amount of periods required for a row to be included
in the result set.
:return: DataFrame with the rolling columns
:raises InvalidPostProcessingError: If the request in incorrect
"""
rolling_type_options = rolling_type_options or {}
df_rolling = df.loc[:, columns.keys()]
kwargs: Dict[str, Union[str, int]] = {}
if window is None:
raise InvalidPostProcessingError(_("Undefined window for rolling operation"))
if window == 0:
raise InvalidPostProcessingError(_("Window must be > 0"))
kwargs["window"] = window
if min_periods is not None:
kwargs["min_periods"] = min_periods
if center is not None:
kwargs["center"] = center
if win_type is not None:
kwargs["win_type"] = win_type
df_rolling = df_rolling.rolling(**kwargs)
if rolling_type not in DENYLIST_ROLLING_FUNCTIONS or not hasattr(
df_rolling, rolling_type
):
raise InvalidPostProcessingError(
_("Invalid rolling_type: %(type)s", type=rolling_type)
)
try:
df_rolling = getattr(df_rolling, rolling_type)(**rolling_type_options)
except TypeError as ex:
raise InvalidPostProcessingError(
_(
"Invalid options for %(rolling_type)s: %(options)s",
rolling_type=rolling_type,
options=rolling_type_options,
)
) from ex
df_rolling = _append_columns(df, df_rolling, columns)
if min_periods:
df_rolling = df_rolling[min_periods:]
return df_rolling
