Handling of NaN values in metric aggregations

[GPalaciosFdez]

Describe the bug
Not really sure if a bug or a design decision but I could not find this case covered in the tests.

When using numpy to calculate the different operations in get_metric_aggregation the functions chosen will return NaN if there is a value in the metric that is NaN.

For example for:

• sum there is nansum - Treats NaN values as 0
• max there is nanmax - Ignores NaN values
• min there is nanmin - Ignores NaN values
• average there is nanmean - Ignores NaN values (Since there are no weights passed to average it behaves the same as mean)
• percentile there is nanpercentile - Ignores NaN values
• std there is nanstd - Ignores NaN values
• var there is nanvar - Ignores NaN values

To Reproduce
Steps to reproduce the behavior:

1. Pass a metric with a NaN value and several numerical values to get_metric_aggregation
2. Check the result and you should see that it is NaN

Expected behavior
Handle NaN values to return a statistically meaningful number

Additional context
I could submit a PR if you think this is something that can be changed.
It could even be possible to pass a boolean flag handle_nans as an argument to get_metric_aggregation to choose the behavior.

[GPalaciosFdez]

Hi :)

Any comments about this issue?

[4n4nd]

@GPalaciosFdez afaik prometheus doesn't really return any nill/nan values. So if there are missing values, we are essentially letting the user manage how the dataframe is created.

Is there a valid case where you see many nans?