Replace workaround by a new xarray lower bound

atlite/datasets/era5.py

@@ -12,7 +12,6 @@
 import os
 import warnings
 import weakref
-from importlib.metadata import version
 from tempfile import mkstemp
 
 import cdsapi
@@ -22,7 +21,6 @@
 from dask import compute, delayed
 from dask.array import arctan2, sqrt
 from numpy import atleast_1d
-from packaging.version import parse
 
 from atlite.gis import maybe_swap_spatial_dims
 from atlite.pv.solar_position import SolarPosition
@@ -363,17 +361,6 @@ def retrieve_data(product, chunks=None, tmpdir=None, lock=None, **updates):
         logger.debug(f"Adding finalizer for {target}")
         weakref.finalize(ds._file_obj._manager, noisy_unlink, target)
 
-    # Remove default encoding we get from CDSAPI, which can lead to NaN values after loading with subsequent
-    # saving due to how xarray handles netcdf compression (only float encoded as short int seem affected)
-    # Fixes issue by keeping "float32" encoded as "float32" instead of internally saving as "short int", see:
-    # https://stackoverflow.com/questions/75755441/why-does-saving-to-netcdf-without-encoding-change-some-values-to-nan
-    # see https://github.com/pydata/xarray/issues/7691 and https://github.com/pydata/xarray/pull/8713
-    # Fix was included in v2024.03.0
-    if parse(version("xarray")) < parse("2024.03.0"):
-        for v in ds.data_vars:
-            if ds[v].encoding["dtype"] == "int16":
-                ds[v].encoding.clear()
-
     return ds
 
 

pyproject.toml

@@ -33,7 +33,7 @@ dependencies = [
         "pandas>=0.25",
         "bottleneck",
         "numexpr",
-        "xarray>=0.20",
+        "xarray>=2024.03.0",
         "netcdf4",
         "dask>=2021.10.0",
         "toolz",
@@ -116,4 +116,4 @@ ignore = [
     'D415', # First line should end with a period, question mark, or exclamation point
     'D417', # Missing argument descriptions in the docstring
 
-]
+]