尝试使用groupby和apply与自定义函数：

import pandas as pd

def cumsum_reset(group: pd.DataFrame) -> pd.Series:
  threshold = group['threshold'].iloc[0]
  cumsum = 0
  reset_values = []
  for value in group['value']:
    cumsum += value
    if cumsum > threshold:
      cumsum = value
    reset_values.append(cumsum)
  return pd.Series(reset_values, index=group.index)

data = {
  'group': ['A', 'A', 'A', 'A', 'B', 'B', 'B', 'C', 'C'],
  'value': [10, 20, 30, 40, 10, 20, 30, 40, 50],
  'threshold': [50, 50, 50, 50, 70, 70, 70, 100, 100]
}
df = pd.DataFrame(data)
df['cumsum_reset'] = df.groupby('group', group_keys=False).apply(lambda g: cumsum_reset(g))
print(df)

字符串

输出：

group  value  threshold  cumsum_reset
0     A     10         50            10
1     A     20         50            30
2     A     30         50            30
3     A     40         50            40
4     B     10         70            10
5     B     20         70            30
6     B     30         70            60
7     C     40        100            40
8     C     50        100            90

型

我建议使用numba来加速计算，例如：

import pandas as pd
from numba import njit

data = {
    "group": ["A", "A", "A", "A", "B", "B", "B", "C", "C"],
    "value": [10, 20, 30, 40, 10, 20, 30, 40, 50],
    "threshold": [50, 50, 50, 50, 70, 70, 70, 100, 100],
}
df = pd.DataFrame(data)

@njit
def get_cumsum_reset_numba(values, threshold, out):
    current = 0
    for i, v in enumerate(values):
        if current + v > threshold:
            out[i] = current
            current = 0
        else:
            current += v
            out[i] = current

def get_cumsum_reset(g):
    g["cumsum_reset"] = 0
    values = g["value"].to_numpy()
    threshold = g["threshold"].iat[0]
    out = g["cumsum_reset"].to_numpy()
    get_cumsum_reset_numba(values, threshold, out)
    return g

df = df.groupby("group", group_keys=False).apply(get_cumsum_reset)
print(df)

字符串
打印：

group  value  threshold  cumsum_reset
0     A     10         50            10
1     A     20         50            30
2     A     30         50            30
3     A     40         50            40
4     B     10         70            10
5     B     20         70            30
6     B     30         70            60
7     C     40        100            40
8     C     50        100            90

型

两件事：

• pandas apply是very slow and should be avoided whenever possible。这在这里是可能的。
• Groupby在处理新组时自然会导致cumsum的重置。因此，这只是一个分组问题，不仅是通过'group'，而且还通过超过阈值。

下面是一个不使用apply的方法：

# 1. Temporary column: first cumsum, grouped by 'group' only
df['cumsum_group'] = df.groupby(by='group')['value'].cumsum()

# 2. Temporary column: times the sum went over the threshold
df['dividend']     = df['cumsum_group'] // df['threshold']

# 3. Final cumsum by double groupby
df['cumsum_reset'] = df.groupby(by=['group','dividend'])['value'].cumsum()
df

  group  value  threshold  cumsum_group  dividend  cumsum_reset
0     A     10         50            10         0            10
1     A     20         50            30         0            30
2     A     30         50            60         1            30
3     A     40         50           100         2            40
4     B     10         70            10         0            10
5     B     20         70            30         0            30
6     B     30         70            60         0            60
7     C     40        100            40         0            40
8     C     50        100            90         0            90

字符串
最后删除临时列：

df.drop(columns=['cumsum_group','dividend'], inplace=True)

型
所有这些都可以总结在这个等效的一行代码中：

df['cumsum_reset'] = df.groupby(by=[df['group'], 
                                    df.groupby(by='group')['value'].cumsum() // df['threshold']]
                               )['value'].cumsum()

型