Assignment

[leezu]

Assignment of individual elements / subsets seems to be only partially supported, but I have not found clear documentation on it. After playing a bit around with xarray I believe the current support boils down to the following (I list some issues I see with the current documentation and how things seem to work).

DataSet

• Change a element in a associated DataArray
  • ds.loc[selector_dict].NAME_OF_VARIABLE = -1 Doesn't work but at least throws an error (Use __setitem__ style assignment (e.g., ds['name'] = ...) instead to assign variables.)
  • ds.loc[selector_dict]["NAME_OF_VARIABLE"] = -1 Silently fails
    • Is it possible to throw an exception as well? I believe this usage variant is "more common", as NAME_OF_VARIABLE may only be known at runtime. Indexing with alignment and broadcasting #974 May fix this (?).
  • Same goes for ds.sel

DataArray
(Can be obtained via ds['name'])

• Using .sel (which is documented to return a new DataArray, so arguably it's expected not to work)
  • ds["NAME_OF_VARIABLE"].sel(**selector_dict) = -1 Invalid python syntax
  • ds["NAME_OF_VARIABLE"].sel(**selector_dict).values = -1 Silently fails
• Using .loc
  • ds["NAME_OF_VARIABLE"].loc[selector_dict] = -1 works

Is this the recommended behaviour? Then I believe the wrong pattern of ds.loc[selector_dict]["NAME_OF_VARIABLE"] = -1 should be documented and ds[NAME_OF_VARIABLE].loc[selector_dict] = -1 instead suggested.

[klapo]

I'm re-pinging this issue since I was just bitten by it (on version 0.15).

If creating an exception is time consuming could we instead have a disclaimer regarding this behavior to the documentation? Here's a basic working example that shows that it does silently fail:

# Basic working example of the silently failing indexing assignment
ds = xr.Dataset({'a': (('x'), np.arange(4)),
                             'b': (('x'), np.arange(4))
                            })
# Silently fails assignment
ds.loc[{'x': 1}]['a'] = 10
print(ds)

# Works as intended
ds['a'].loc[{'x': 1}] = 10
print(ds)

[keewis]

it's not that it fails, the original dataset is just not updated:

In [2]: ds = xr.Dataset({ 
   ...:     "a": ("x", np.arange(4)), 
   ...:     "b": ("x", np.arange(4)), 
   ...: })

In [3]: selected = ds.loc[{"x": 1}]
   ...: selected["a"] = 10
   ...: ds, selected
Out[3]:
(<xarray.Dataset>
 Dimensions:  (x: 4)
 Dimensions without coordinates: x
 Data variables:
     a        (x) int64 0 1 2 3
     b        (x) int64 0 1 2 3,
 <xarray.Dataset>
 Dimensions:  ()
 Data variables:
     a        int64 10
     b        int64 1)

to make it work, use (note the [...])

In [7]: selected = ds.loc[{"x": 1}] 
   ...: selected["a"][...] = 10 
   ...: ds, selected
Out[7]: 
(<xarray.Dataset>
 Dimensions:  (x: 4)
 Dimensions without coordinates: x
 Data variables:
     a        (x) int64 0 10 2 3
     b        (x) int64 0 1 2 3,
 <xarray.Dataset>
 Dimensions:  ()
 Data variables:
     a        int64 10
     b        int64 1)

The difference is that without the extra [...] we don't modify the values but instead create a new variable (so something like ds["new_variable"] = ("x", [1, 3, 4, 5]))

Also note that while ds.a.sel(x=1) = 10 raises a SyntaxError, ds.a.sel(x=1)[...] = 10 or ds.sel(x=1).a[...] = 10 have exactly the same effect as their loc counterparts.

We probably should add an explanation and examples to the documentation (loc / sel and maybe the indexing page?)

[keewis]

also, we can't really raise only for ds.loc[...][name] = -1 since for python that is the same as

intermediate_ds = ds.loc[...]
intermediate_ds[name] = -1

where the second line is something we explicitly want to support.

[keewis]

since it came up in #3939, I played around with this, and it seems the obj.sel(x=1)[...] = 0 syntax only works for DataArray objects and when assigning a single value. Should we try to make it work the same way as loc?

[dcherian]

How does this even work? 😱

[stale]

In order to maintain a list of currently relevant issues, we mark issues as stale after a period of inactivity

If this issue remains relevant, please comment here or remove the stale label; otherwise it will be marked as closed automatically