@Huite Indeed, you are right that working with a coordinate is easy if it works for DataArrays ... this is a good example of my pandas-oriented brain not quite being used to xarray just yet (though I do love it).

Regarding signature options for a Dataset ... given the two examples you state, I also personally prefer the look of the second one. However, the first one can be extremely useful for more complicated replacement needs because the input dict can be assembled programmatically prior to the replace call, for doing replaces in several subset DataArrays. I think the second version would require looping of some sort, or multiple calls at the very least. For me, in my context of renaming on coordinates (the index or columns in a DataFrame context), I often have to modify many things in both axes, which I do using one dictionary.

I suppose it's a matter of preference and of ease of implementation ... since I'm not the one doing the coding, I shall definitely defer to others on the latter point!

@Jeitan

The coordinate is a DataArray as well, so the following would work:

```python

Example DataArray

da = xr.DataArray(np.ones((3, 3)), {"y": [50.0, 60.0, 70.0], "x": [1.0, 2.0, 3.0]}, ("y", "x"))

Replace 50.0 and 60.0 by 5.0 and 6.0 in the y coordinate

da["y"] = da["y"].replace_values([50.0, 60.0], [5.0, 6.0]) ```

Your example in the other issue mentions one of the ways you'd replace in pandas, but for a dataframe. With a dataframe, there's quite some flexibility:

python df.replace({0: 10, 1: 100}) df.replace({'A': 0, 'B': 5}, 100) df.replace({'A': {0: 100, 4: 400}})

I'd say the xarray counterpart of a Dataframe is a Dataset; the counterpart of a DataArray is a Series. Replacing the coordinates in a DataArray is akin to replacing the values of the index of a Series, which is apparently possible with series.rename(index={from: to}).

Other thoughts: some complexity comes in when implementing a replace_values method for a Dataset. I also think the pandas replace method signature is too complicated (scalars, lists, dicts, dicts of dicts, probably more?) and the docstring is quite extensive (https://pandas.pydata.org/docs/reference/api/pandas.DataFrame.replace.html)

I think the question is what the signature should be. You could compare to reindex (https://xarray.pydata.org/en/stable/generated/xarray.Dataset.reindex.html) and have an "replacer" argument:

```python da = da.replace({"y": ([50.0, 60.0], [5.0, 6.0])})

da["y"] = da["y"].replace([50.0, 60.0], [5.0, 6.0]) ```

The first one would also work for Datasets, but I personally prefer the second one for it's simplicity (and which is maybe closer to .where : https://xarray.pydata.org/en/stable/generated/xarray.DataArray.where.html).

Thanks @dcherian for linking the other issue because that led me here. I'm all for this! Though I would like to add the consideration for doing this replacement in a coordinate, not just the data (parts of the suggested code like returning da.copy(data=out.reshape(da.shape)) won't work for that). Once they are accessed coordinates work very much like the data part, so hopefully making this general shouldn't be too hard?

Nice find @dcherian .

So it sounds like there's consensus around something like replace_data / replace_values / update_values. If you'd still be up for putting together a PR, I think that would be very welcome.

You're right about np.select @Huite . The np.searchsorted solution looks v clever!

See also #5048 though the discussion here is more thorough.

Yeah I think maybe replace_values is better name. "search and replace values" is maybe how you'd describe it colloquially?remap is an option too, but I think many users won't have the right assocation with it (if they're coming from a less technical background).

I don't think you'd want to this with np.select. If I understand correctly, you'd have to broadcast for the number of values to replace. This work okay with a small number of replacement values, but not with 10 000 like in my example above (but my understanding might be lacking).

Having said that, there is a faster and much cleaner implementation using np.seachsorted on da instead.

```python def custom_replace2(da, to_replace, value): flat = da.values.ravel()

sorter = np.argsort(to_replace)
insertion = np.searchsorted(to_replace, flat, sorter=sorter)
indices = np.take(sorter, insertion, mode="clip")
replaceable = (to_replace[indices] == flat)

out = flat.copy()
out[replaceable] = value[indices[replaceable]]
return da.copy(data=out.reshape(da.shape))

For small example: 4.1 ms ± 144 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

For the larger example: # 14.4 ms ± 592 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

%timeit custom_replace2(da, to_replace, value) ```

This is equal to the implementation of remap in numpy-indexed (which is MIT-licensed): https://github.com/EelcoHoogendoorn/Numpy_arraysetops_EP

The key trick is the same, relying on sorting.

See e.g. also: https://stackoverflow.com/questions/16992713/translate-every-element-in-numpy-array-according-to-key

I agree this would be useful, and I've had to do similar things. It's the sort of area where pandas is stronger than xarray.

We might want a more specific name than replace; something that confers it's replacing values? Particularly if the method is on a Dataset as well as a DataArray.

@Huite thanks for the great proposal. Did you look at np.select? I think that might be faster than these and require less code.