I’m trying to convert the following R code to python

stdev.cust <- function(x){
 sd(x, na.rm = TRUE)
}

.stdev.min.obs <- function(x, min.n) {
    n.s <- sum(!is.na(x))
    ifelse(n.s >= min.n, stdev.cust(x), NA_real_)
}


df[, col2  := rollapply(col1,
                        width=35,
                        .stdev.min.obs,
                        min.n=10,
                        fill=NA_real_,
                        partial=TRUE,
                        align='right'),
  keyby='date']

This is what I came up with

def stdev_min_obs(x, min_n):
    ns = np.sum(~np.isnan(x))
    return np.std(x) if ns >= min_n else np.nan

df['col2'] = df.groupby('date')['col1'].apply(
    lambda x: x.rolling(window=35, min_periods=10).apply(lambda y: stdev_min_obs(y, 10))
)

But when I tried this on the problem I’m working on at work, the results aren’t the same, and also the python version is a lot slower, presumably because apply is very slow?