issue_comments

@JoerivanEngelen Thanks for taking the time. Much appreciated.

@kmuehlbauer your fix fixes both the problems set up by @slevang and @amatsukawa on my laptop.

I furthermore tested the double roundtrip with load_dataset and to_netcdf, with the three possible engines, they all worked.

Please check #7720 if that fixes the conversion problems. Thanks.

Closed https://github.com/pydata/xarray/issues/5192 in favor of this as I think it's a duplicate. Just NB that it can occur with h5netcdf as well as netcdf4. (Thanks, @andersy005 )

I think this issue is related to #2937.

This is the same problem reported in https://github.com/pydata/xarray/issues/4386. All these issues (#2937, #4826, #4386, #5192) should be closable once there's a fix.

There are at least two issues here:

• For both netCDF and zarr, we should definitely "round trip" preserve dtypes. There are a few ways this might be done. I would have to look at a PR to say for sure, but from the looks of it "Change maybe_encode_bool to not ignore when .encoding["dtype"] exists" is the right fix.
• Zarr supports bool data, so we don't need conversion from bool -> int8. We should just save data as bool.

I ran into this as well with the basic netcdf backends:

```python import xarray as xr

ds = xr.Dataset( data_vars={"foo":(["x"], [False, True, False])}, coords={"x": [1, 2, 3]}, ) ds.to_netcdf('test.nc') ds = xr.load_dataset('test.nc') print(ds.foo.dtype) ds.to_netcdf('test.nc') ds = xr.load_dataset('test.nc') print(ds.foo.dtype) Gives: bool int8 ```

Tagging a few maintainers: @dcherian @shoyer.

Sorry to tag you directly, hope that's ok. I think I've found the issue here and would like to provide a PR to fix, but need some input on what you think would be best.

To summarize, the current behavior leading to the bug is:

1. When a bool dtype is initially written, the maybe_encode_bool function us used to convert it the bool to a i1 with a vars.attr that says it is actually a bool. It would appear from #2937 that it ends up an i8 somehow anyway. https://github.com/pydata/xarray/blob/cc53a77ff0c8aaf8686f0b0bd7f75985b74e2054/xarray/conventions.py#L119
2. When this is loaded the first time, the i8 is correctly identified as actually being a bool using the attributes. https://github.com/pydata/xarray/blob/cc53a77ff0c8aaf8686f0b0bd7f75985b74e2054/xarray/conventions.py#L352
3. However, 2 lines above that, there is a encoding.setdefault("dtype", original_dtype) so this Variable object now has .encoding["dtype"] which is i8.
4. When I try to save this again, it tries maybe_encode_bool from step 1 again, but this time the function is bypassed because of step 3 above.
5. The dataset I write from step 4 now does not have the attribute identifying it as a bool, and so it's an i8 when I load it back.

I can think of a few fixes: - Drop the .encoding dict on load for bools. Presumably these .encodings are kept such that datasets attempts to preserve the compressor, chunk size, etc. of its source. However, given that .encoding seems to be dropped when I eg. do a .astype("bool") maybe this is OK for bools. - Set var.attrs["dtype"] = np.dtype("bool") on load. This would preserve what we'd get out of maybe_encode_bool. The challenge with this is that .attrs are not always preserved? - Change maybe_encode_bool to not ignore when .encoding["dtype"] exists. I suppose there would be a need to check that the compressor is still compatible, etc?

As a local fix while we consider these options, can you confirm that, as the docs state, the .encoding is only used for serializing and not deserializing arrays, and therefore if I drop the .encoding on my DataArrays as a temporary fix I wouldn't break anything (if I am ok with xarray not preserving my compressors, etc).

OK here's the other side of the problem. The original dtype (which is i8) is set in the encoding:

https://github.com/pydata/xarray/blob/v0.16.2/xarray/conventions.py#L350

Apparently my proposed fix broke a bunch of other things, eg. some writing of timedeltas with units and such.

Deleting the "dtype" key in the .encoding of the boolean variable also seems to do the trick. The issue is that bools are not encoded correctly if the .encoding field already has a dtype:

https://github.com/pydata/xarray/blob/v0.16.2/xarray/conventions.py#L119