Alright, I think I get it, thanks for the clarification @crusaderky

Thing is, the whole thing is undefined. What does the accessor state contain? As a xarray developer, I don't know. Is it variable names? Is it references to objects that make up the Dataset, e.g. Variables or the attrs dict? Is it objects whose contents rely on the current state of the Dataset, e.g. aggregated measures? Is it objects whose contents rely on historical events (like in your case)?

Dataset.copy() will create a copy of everything up to and excluding the numpy arrays. In order to allow you to retain accessor state, we'd need to plant a hook in it and invoke some agreed duck-type API in your object that basically states, "I called copy(), and this is the new object I created, please create a copy of yourself accordingly making extra sure you don't retain references to components of the previous object". And then there are all the other methods that currently nuke the accessor state - including many in-place ones - because they could potentially invalidate it. What should they do? Invoke a special API on the accessor? If not, why should copy() trigger special accessor API and e.g. roll() shouldn't? Planting accessor-refresher hooks in every single method that currently just wipes it away is out of question as it would need to be almost everywhere and - more importantly - it would be born broken.

Ok, I get it. Probably the accessor is not the best solution in my case. And yes, an attribute was in fact my first implementation of the add/clean idea. But I was afraid it would be less reliable than the internal list over a long term perspective (but that was before getting in the troubles described above).

But why is asking accessor developers to define a copy method an issue ? That wouldn't be mandatory but only required in situations where propagating functional informations could be useful. Sorry if that's a naive question for you guys.

Would that make any sense that the xr.DataSet.copy() method also return a copy of the accessors ?

It's been discussed above in this same thread. It's impossible without breaking the accessor API, as it would require you (the accessor developer) to define a copy method. The more high level discussion is that the statefulness of the accessor is something that is OK to use for caching and performance improvements, and not OK for storing functional information like yours.

Have you considered storing a flag in Variable.attrs instead?

```python

def add(self, da):
    da.attrs["cleanable"] = True
    self.obj[da.name] = da
    return self.obj

def clean(self):
    return self.obj.drop([
        k for k, v in self.obj.variables.items()
        if v.attrs.get("cleanable")
    ])

```

@crusaderky thanks for the explanation, that's a solution to my pb.

Although I understand that since accessor will be created from scratch, a dataset copy won't propagate the accessor properties (in this case the list of added variables):

```python ds = xarray.Dataset() ds['ext_data'] = xarray.DataArray(1.)

my_estimator = BaseEstimator() # With "clean" method from @crusaderky ds.my_accessor.fit(my_estimator, x=2.) ds.my_accessor.transform(my_estimator, y=3.)

ds2 = ds.copy()

ds = ds.my_accessor.clean() ds2 = ds2.my_accessor.clean()

print(ds.data_vars) print(ds2.data_vars) gives:python Data variables: ext_data float64 1.0 Data variables: ext_data float64 1.0 fit_data float64 4.0 trf_data float64 7.0 ``` "Cleaning" the dataset works as expected, but the copy (ds2) has en empty list of added variables so the "clean" method doesn't have the expected result. We have the same behavior for deep copy.

Would that make any sense that the xr.DataSet.copy() method also return a copy of the accessors ?

@gmaze Dataset.drop does not mutate the state of the original object, so it's conceptually wrong for your clean() method to mutate the accessor state too. It should be: python def clean(self): return self.obj.drop(self.added) The new dataset returned will have no accessor cache, and will recreate an instance on the fly on first access.

Hi all, I recently encountered an issue that look like this with accessor, but not sure. Here is a peace of code that reproduces the issue.

Starting from a class with the core of the code and an accessor to implement the user API:

``` python import xarray

class BaseEstimator(): def fit(self, this_ds, x=None): # Do something with this_ds: x = x**2 # and create a new array with results: da = xarray.DataArray(x).rename('fit_data') # Return results: return da

def transform(self, this_ds, **kw):
    # Do something with this_ds:
    val = kw['y'] + this_ds['fit_data']
    # and create a new array with results:
    da = xarray.DataArray(val).rename('trf_data')
    # Return results:
    return da

@xarray.register_dataset_accessor('my_accessor') class Foo: def init(self, obj): self.obj = obj self.added = list()

def add(self, da):
    self.obj[da.name] = da
    self.added.append(da.name)
    return self.obj

def clean(self):
    for v in self.added:
        self.obj = self.obj.drop(v)
        self.added.remove(v)
    return self.obj

def fit(self, estimator, **kw):
    this_da = estimator.fit(self, **kw)
    return self.add(this_da)

def transform(self, estimator, **kw):
    this_da = estimator.transform(self.obj, **kw)
    return self.add(this_da)

```

Now if we consider this workflow: ``` python

ds = xarray.Dataset() ds['ext_data'] = xarray.DataArray(1.)

my_estimator = BaseEstimator() ds = ds.my_accessor.fit(my_estimator, x=2.)

print("Before clean:") print("xr.DataSet var :", list(ds.data_vars)) print("accessor.obj var:", list(ds.my_accessor.obj.data_vars))

print("\nAfter clean:")

ds.my_accessor.clean() # This does nothing to ds but clean the accessor.obj

ds = ds.my_accessor.clean() # Cleaning ok for both ds and accessor.obj

ds_clean = ds.my_accessor.clean() # Cleaning ok on new ds, does nothing to ds as expected but clean in accessor.obj print("xr.DataSet var :", list(ds.data_vars)) print("accessor.obj var :", list(ds.my_accessor.obj.data_vars)) print("Cleaned xr.DataSet var:", list(ds_clean.data_vars)) We have the following output:python Before clean: xr.DataSet var : ['ext_data', 'fit_data'] accessor.obj var: ['ext_data', 'fit_data']

After clean: xr.DataSet var : ['ext_data', 'fit_data'] accessor.obj var : ['ext_data'] Cleaned xr.DataSet var: ['ext_data'] ``` The issue is clear here: the base space dataset has the 'fit_data' variable but not the accessor object: they've been "disconnected" and it's not apparent to users.

So if users later proceed to run the "transform":

python ds.my_accessor.transform(my_estimator, y=2.) they get an KeyError raised because the 'fit_data' is not in the accessor, although it still appears on the list of the ds variables, which is more than confusing.

Sorry for this long post, I'm not sure it's relevant to this issue but it seems so to me. I don't see a solution to this from the accessor developer side, except for not "interfering" with the content of the accessed object.

I don't mind needing to update our accessor code. My only request is don't have a version that suddenly breaks it so that we only work on the old version or the new version. 😉

It isn't just methods that use _to_temp_dataset() that result in losing the accessor state -- any operation that creates a new DataArray will (by design) lose the accessor, which don't get propagated in any operations.

Accessors are also not preserved when indexing a DataArray out of a Dataset (https://github.com/pydata/xarray/issues/3205), cc @djhoese.

I had not contemplated the issue of circular references, which I agree is not ideal. If we had realized that when creating accessors in the first place we might have chosen a different design, but there are a number of users who rely upon it.

The circular reference issue could also be worked around in a user-friendly way by having the decorator automatically add methods to the decorated class

I like the look of this solution. It will still require users to update their code to avoid circular references (e.g., by removing their own __init__ method), but it will make the default behavior more sane.

could you change your accessor code to store its state in Dataset.attrs instead?

Yes, although attrs would be bad as well because they are lost in many operations. The current "best practice" would be to use coords for this, as in https://github.com/pydata/xarray/issues/1271#issuecomment-280574680

I never bothered to get this done (my library is quite niche and not mean to become widely used), but if we remove the caching mechanism in xarray, it would be one more incentive to make the switch.

So you work on the assumption that no new potentially useful coords will be added after the first invocation of your accessor?

Yes. In almost all use cases I know of this won't happen. Typically, the users have to create (or open) an xarray object that salem understands before calling the accessor anyway.

Or do you have logic that invalidates your cache every time the state of the coords changes?

No. I didn't think about that before today ;-). But you are right: if users of salem's accessor change the coordinates after calling the accessor, then bad things might happen. Experience shows that this rarely happens (never happened to me), but I can see how this can fire back.

Altogether, these are good arguments to remove the caching mechanism I believe.

The circular reference issue could also be worked around in a user-friendly way by having the decorator automatically add methods to the decorated class, copying the design of @dataclass:

```python import weakref

class C: def init(self, owner): self._owner = weakref.ref(owner) if hasattr(self, "post_init"): self.post_init()

@property
def owner(self):
    out = self._owner()
    if out is None:
        raise AttributeError("Orphaned accessor")
    return out

``` This would also allow for shallow copies to change the pointer.

store the names of the coordinate variabless we know are going to be useful to us later. So you work on the assumption that no new potentially useful coords will be added after the first invocation of your accessor? Or do you have logic that invalidates your cache every time the state of the coords changes?

@fmaussion could you change your accessor code to store its state in Dataset.attrs instead?

Interesting, thanks!

As an accessor maintainer, I can ensure that at least one accessor implementation is storing state ;-). But this state is based on the xarray object itself: for example, we derive georeferencing information and store the names of the coordinate variabless we know are going to be useful to us later. That is, every new call to __init__ based on a modified object will trigger a new parsing, and we don't come into the situation you describe above.

Getting rid of the caching logic would mean some performance loss to us, yes, but I don't know if it's "worse" than the circular reference issue you describe or not.

Demonstation on the circular reference issue: ```python import gc import weakref import xarray

class C: pass

@xarray.register_dataset_accessor('foo') class Foo: def init(self, obj): self.obj = obj

ds = xarray.Dataset() w = weakref.ref(ds) print("No accessor, in scope:", w() is not None) del ds print("No accessor, descoped:", w() is not None)

ds = xarray.Dataset() ds.foo w = weakref.ref(ds) print("with accessor, in scope:", w() is not None) del ds print("with accessor, descoped:", w() is not None) gc.collect() print("with accessor, after gc pass:", w() is not None) Output: No accessor, in scope: True No accessor, descoped: False with accessor, in scope: True with accessor, descoped: True with accessor, after gc pass: False ```