Newer
Older
Matthew K Defenderfer
committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
import cudf
import pandas as pd
import dask.dataframe as dd
import dask_cudf
from .utils import as_timedelta
from typing import Literal
from typeguard import typechecked
__all__ = ['get_aggregator']
# ENH: In the future, probably need to wrap the manager and backend type into a class. That class would contain the
# read_parquet function instead of putting it in the aggregation classes. This would separate everything out more
# sensibly
@typechecked
class Aggregator:
def __init__(self):
self.backend = None
self.cuda = None
def _cut(
self,
ser: pd.Series | cudf.Series,
bins: list[int | pd.Timestamp],
labels: list[str] | None = None,
**kwargs
) -> pd.Series | cudf.Series:
right = kwargs.pop('right',False)
if self.cuda:
func = cudf.cut
ser = ser.astype('int64')
else:
func = pd.cut
grps = func(ser,bins=bins,labels=labels,right=right,**kwargs)
if labels is not None:
grps = grps.cat.reorder_categories(labels[::-1], ordered = True)
return grps
def create_timedelta_cutoffs(
self,
delta_vals: int | list[int],
delta_unit: Literal['D','W','M','Y'],
run_date: pd.Timestamp
) -> list[int | pd.Timestamp]:
deltas = pd.Series([as_timedelta(c,delta_unit) for c in delta_vals])
cutoffs = pd.to_datetime(run_date - deltas)
cutoffs = (
pd.concat(
[
cutoffs,
pd.Series([pd.to_datetime('today'),pd.to_datetime('1970-01-01')])
]
)
.sort_values()
)
return cutoffs.astype('int64').to_list() if self.cuda else cutoffs.to_list()
def create_timedelta_labels(
self,
delta_vals: list[int],
delta_unit: Literal['D','W','M','Y'],
) -> list[str]:
delta_vals.sort(reverse=True)
deltas = [f'{d}{delta_unit}' for d in delta_vals]
labels = [f'>{deltas[0]}'] + [f'{deltas[i+1]}-{deltas[i]}' for i in range(len(deltas)-1)] + [f'<{deltas[-1]}']
return labels
class PandasAggregator(Aggregator):
def __init__(self):
self.backend = 'pandas'
self.cuda = False
def read_parquet(self,dataset_path,**kwargs) -> pd.DataFrame:
return pd.read_parquet(dataset_path,**kwargs)
def cut_dt(self,series,*args,**kwargs) -> pd.Series:
return self._cut(series,*args,**kwargs)
def aggregate(
self,
df: cudf.DataFrame,
col: str | list[str],
grps: str | list[str],
funcs: str | list[str]
) -> pd.DataFrame:
df_agg = (
df.groupby(grps,observed = True)[col]
.agg(funcs)
.sort_index(level=[0,1])
.reset_index()
)
return df_agg
class CUDFAggregator(Aggregator):
def __init__(self):
self.backend = 'cudf'
self.cuda = True
def read_parquet(self,dataset_path,**kwargs) -> cudf.DataFrame:
return cudf.read_parquet(dataset_path,**kwargs)
def cut_dt(self,series,*args,**kwargs) -> pd.Series:
return self._cut(series,*args,**kwargs)
def aggregate(
self,
df: cudf.DataFrame,
col: str | list[str],
grps: str | list[str],
funcs: str | list[str]
) -> pd.DataFrame:
df_agg = (
df.groupby(grps,observed = True)[col]
.agg(funcs)
.sort_index(level=[0,1])
.to_pandas()
.reset_index()
)
return df_agg
class DaskAggregator(Aggregator):
def __init__(self):
self.backend = 'dask'
self.cuda = False
def cut_dt(self,series,*args,**kwargs) -> cudf.Series:
return series.map_partitions(self._cut,*args,**kwargs)
def aggregate(
self,
df: dd.DataFrame,
col: str | list[str],
grps: str | list[str],
funcs: str | list[str]
) -> pd.DataFrame:
df_agg = (
df.groupby(grps,observed = True)[col]
.agg(funcs)
.compute()
.sort_index(level=[0,1])
.reset_index()
)
return df_agg
def read_parquet(self,dataset_path,**kwargs) -> dd.DataFrame:
split_row_groups = kwargs.pop('split_row_groups',False)
return dd.read_parquet(dataset_path,split_row_groups=split_row_groups,**kwargs)
class DaskCUDFAggregator(Aggregator):
def __init__(self):
self.backend = 'dask_cuda'
self.cuda = True
def cut_dt(self,series,*args,**kwargs) -> dask_cudf.Series:
return series.map_partitions(self._cut,*args,**kwargs)
def aggregate(
self,
df: dask_cudf.DataFrame,
col: str | list[str],
grps: str | list[str],
funcs: str | list[str]
) -> pd.DataFrame:
df_agg = (
df.groupby(grps,observed = True)[col]
.agg(funcs)
.compute()
.sort_index(level=[0,1])
.to_pandas()
.reset_index()
)
return df_agg
def read_parquet(self,dataset_path,**kwargs) -> dd.DataFrame:
split_row_groups = kwargs.pop('split_row_groups',False)
return dd.read_parquet(dataset_path,split_row_groups=split_row_groups,**kwargs)
def get_aggregator(backend) -> PandasAggregator | CUDFAggregator | DaskAggregator | DaskCUDFAggregator:
match backend:
case 'pandas':
return PandasAggregator()
case 'cudf':
return CUDFAggregator()
case 'dask':
return DaskAggregator()
case 'dask_cuda':
return DaskCUDFAggregator()
case _:
raise ValueError(f"Unsupported backend: {backend}")