Milan Klöwer

> What I do not want is that the compression impacts the read performance too much, i.e. reading the compressed data should not take much more (subjective I know) time...

> where is the best place to discuss usage/interpretation/best practices with usage of these tools? Here 😄 > I'm wondering what it means that some variables don't have continuous distributions...

Regarding SST, and other data that will include some form of missing values, this is something that provided me a lot of headache in the past. Reason being that you...

Yes, you can do that. However, it might an overkill. I found the bitwise information fairly robust in general, such that I'd probably start by looking at a time step...

> The bitinformation method is applicable for relative precision trimming [...] the signal must have purely multiplicative noise This is correct when used with floats, but due to floats and...

Happy to keep this open, if there's more on a given topic that risks other points being lost here feel free to branch off into another issue.

> I am not sure if "O3 mixing ratio field keeping 99% of information" (or should one keep 99.9%?) is any more defensible than 'accurate within 1/128 or 0.0625ppb'. At...

In principle I agree with the argument that one should consider different uncertainty quantifications if possible. But it's rarely straightforward what the uncertainty of a dataset is. Sure, one may...

So on the note of best practices, I would say - use the bitwise information analysis to determine a number of keepbits - check whether the precision/size trade-off works well...

Hi @aaronspring many thanks for sharing this analysis! 1) Could you specify how your data is structured? I see that you are using some less regular grid (is that tripolar?)...