Faster rolling apply std and slope on unevenly-spaced timeseries - python-3.x

Question

I have a Pandas (1.0.*) dataframe, which contains the record of several physical variables (say Temperature, Pressure and Humidity for example).
The time space between two record is roughly 1s but varies between 0.8s and 4s.
I want to calculate the standard deviation and the slope (of the linear regression) in a 5-minutes rolling window.
Here is how I do it:
import numpy as np
import pandas as pd
import datetime
np.random.seed(1)
# Build the dummy dataset for testing
rows, cols = 1000, 3
datetimes_sec = pd.date_range('2020-01-01', periods=rows, freq='1s').astype(np.int64) / 1e9
shifts = np.random.rand(rows) - 0.5 # Create random shift between -0.5s and +0.5s
datetimes = [sum(x) * 1e9 for x in zip(datetimes_sec, shifts)]
df = pd.DataFrame(np.random.rand(rows,cols),
columns=['temperature', 'pressure', 'humidity'],
index=pd.to_datetime(datetimes))
# Custom function to calculate the slope
def get_slope(series):
hours_since_epoch = series.index.astype(np.int64) / 3.6e12 # nanosecond to hour, I want the slope to be in [variable's unit] per hour
slope = np.polyfit(hours_since_epoch, series, 1)[0]
return slope
# Get the result
df = df.rolling("5min").agg(["std", get_slope])
This works, but it is too slow: the last line takes more than 2s for 1000 rows.
I can see that my custom get_slope function is responsible, if I replace it by a standard function (e.g. min()), it takes 0.007s. But I can find how to make it faster.
If it is not possible to get the same result faster, a workaround solution could be to skip some data lines: do not roll the window on every line (i.e. for 0.8 to 4 second) but do the calculation only every 30s:
calculate sd and slope of all (~300) data between 00:00:00 and 00:05:00
calculate sd and slope of all (~300) data between 00:00:30 and 00:05:30
calculate sd and slope of all (~300) data between 00:01:00 and 00:06:00
etc.
Instead of:
calculate sd and slope of all (~300) data between 00:00:00 and 00:05:00
calculate sd and slope of all (~300) data between 00:00:01 and 00:05:01
calculate sd and slope of all (~300) data between 00:00:02 and 00:05:02
etc.
I don't know how to do that (in the proper pandas' way) with unevenly-spaced data.
It would speed up the process by 30, in exchange of a loss of precision.