Always force dask arrays to float in missing.interp_func (#4771)

* scipy.interpolate.interp1d always forces to float.

* Copy type-check from scipy.interpolate.interp1d

* Update missing.py

* Test that pre- and post-compute dtypes matches

* Update test_missing.py

xarray/core/missing.py

@@ -730,6 +730,13 @@ def interp_func(var, x, new_x, method, kwargs):
         # if usefull, re-use localize for each chunk of new_x
         localize = (method in ["linear", "nearest"]) and (new_x[0].chunks is not None)
 
+        # scipy.interpolate.interp1d always forces to float.
+        # Use the same check for blockwise as well:
+        if not issubclass(var.dtype.type, np.inexact):
+            dtype = np.float_
+        else:
+            dtype = var.dtype
+
         return da.blockwise(
             _dask_aware_interpnd,
             out_ind,
@@ -738,7 +745,7 @@ def interp_func(var, x, new_x, method, kwargs):
             interp_kwargs=kwargs,
             localize=localize,
             concatenate=True,
-            dtype=var.dtype,
+            dtype=dtype,
             new_axes=new_axes,
         )
 

xarray/tests/test_missing.py

@@ -370,6 +370,20 @@ def test_interpolate_dask_raises_for_invalid_chunk_dim():
         da.interpolate_na("time")
 
 
+@requires_dask
+@requires_scipy
+@pytest.mark.parametrize("dtype, method", [(int, "linear"), (int, "nearest")])
+def test_interpolate_dask_expected_dtype(dtype, method):
+    da = xr.DataArray(
+        data=np.array([0, 1], dtype=dtype),
+        dims=["time"],
+        coords=dict(time=np.array([0, 1])),
+    ).chunk(dict(time=2))
+    da = da.interp(time=np.array([0, 0.5, 1, 2]), method=method)
+
+    assert da.dtype == da.compute().dtype
+
+
 @requires_bottleneck
 def test_ffill():
     da = xr.DataArray(np.array([4, 5, np.nan], dtype=np.float64), dims="x")