hvplot / holoviews (bokeh) timedelta axis with negative values formatter

Question

Issue:

I am trying to plot data with a pandas time delta index with negative time delta values on the x-axis using hvplot or holoviews (bokeh backend).

The labels are just integers, and seem to be in milliseconds.
I want them to be formatted in a friendlier format such as HH:MM:SS

EXAMPLE

import pandas as pd
import numpy as np
import hvplot.pandas

x = pd.timedelta_range(start=0, freq='S', periods=11) - pd.Timedelta('5S')
y = np.arange(len(x))
df = pd.DataFrame({'y': y}, index=x)
df.hvplot.line(rot=20)

Output: I expected the x axis to be -00:00:04 -00:00:04 00:00:00 00:00:02 00:00:04
or at least in seconds, this seems to be milliseconds.

What I tried

using the df create above:

from bokeh.models.formatters import NumeralTickFormatter
df.hvplot.line(xformatter=NumeralTickFormatter(format="00:00:00"), rot=20)

Output: Idk what happened here with the xlabels but they dont really make any sense.

Using DatetimeTickFormatter:

from bokeh.models.formatters import DatetimeTickFormatter
df.hvplot.line(rot=20, xformatter=DatetimeTickFormatter())

Does not work unfortunatelt: No negative values: - 00:00:02 becomes 58s

Answer 1

I found a solution, it is not pretty but it works
(Using the earlier created dataframe):

def timedelta_formatter(x):
    x/=1000                       # ms -> seconds

    # extract seconds, minutes, hours, days from time
    m, s = divmod(abs(x), 60)
    h, m = divmod(m, 60)
    d, h = divmod(h, 24)
    
    # create output string with time format
    out = f"{d}d. {h:02}:{m:02}:{s:02}"

    if x < 0:
        # add - if negative value
        return "- " + out
    else:
        return out
df.hvplot.line(xformatter=timedelta_formatter, rot=20)

Output:

This looks better!

Lets try it with a timedelta range in days:

x = pd.timedelta_range(start=0, freq='D', periods=11) - pd.Timedelta('5D')
y = np.arange(len(x))
df = pd.DataFrame({'y': y}, index=x)
df.hvplot.line(xformatter=timedelta_formatter, rot=20)

Output:

Works too

I still think this is not the most elegant solution,
so open to hear better solutions