Change default fill values

[TomAugspurger]

Zarr version

v3

Numcodecs version

na

Python Version

na

Operating System

na

Installation

na

Description

Over in pydata/xarray#5475, we've been discussing an issue that's affecting xarray with Zarr v3. xarray currently interprets values equal to fill_value as "missing" and casts them to NaN (apparently in NetCDF (or CF conventions?) there's some understanding that its _FillValue is understood to be outside the "valid range" of the data).

There's lots of discussion there, but one thing zarr-python could do to help would be to choose default fill values that are less likely overlap with valid data. Exactly what's valid is domain / application / dataset specific, but I think that 0 (or the equivalent for some dtype) is slightly more likely to be valid than many others, and so might be a worse default.

What do people thing about the following kinds of rules?

• (signed) integer: intmin / np.iinfo(dtype).min
• unsigned integer: intmax / np.iinfo(dtype).max
• float: nan
• complex: nan+nan0j

Steps to reproduce

na

Additional output

No response

[jhamman]

I'm a big +1 on changing the defaults. I don't really care about the values but will note NetCDF4 has also defined defaults:

In [1]: from netCDF4 import default_fillvals

In [2]: default_fillvals
Out[2]:
{'S1': '\x00',
 'i1': -127,
 'u1': 255,
 'i2': -32767,
 'u2': 65535,
 'i4': -2147483647,
 'u4': 4294967295,
 'i8': -9223372036854775806,
 'u8': 18446744073709551614,
 'f4': 9.969209968386869e+36,
 'f8': 9.969209968386869e+36}

[d-v-b]

I'm not opposed to moving the defaults, but it's probably worth hearing from people in other domains. In my experience in bioimaging, for both raw images and segmentations, 0 is conventionally used as a background label and people very often rely on application default values (which is probably where the convention came from in the first place).

cc @jni

Maybe in zarr v4 we can have proper support for nullable types to avoid this problem alltogether ;)