issue_comments

Let me know any post-merge feedback and I'll make the changes

Any final feedback before merging?

it definitely is not: look at the builds of #4809: one build failed and the next (about 12 minutes later) fails. I guess either sphinx, nbsphinx or one of their dependencies got updated on conda-forge. nbformat (5.0.8 → 5.1.0), maybe?

Edit: nbformat=5.1.2 is out which should fix that issue.

Would anyone know whether the docs failure is related to this PR? I can't see anything in the log apart from matplotlib warnings? https://readthedocs.org/projects/xray/builds/12768566/

I double-checked the benchmarks and added a pandas comparison. That involved ensuring the missing value was handled corre them and ensured the setup wasn't in the benchmark.

I don't get the 100x speed-up that I thought I saw initially; it's now more like 8x. Still decent! I'm not sure whether that's because I misread the benchmark previously or because the benchmarks are slightly different — I guess the first.

Pasting below the results so we have something concrete.

Existing asv profile unstacking.Unstacking.time_unstack_slow master | head -n 20 ··· unstacking.Unstacking.time_unstack_slow 861±20ms

Proposed asv profile unstacking.Unstacking.time_unstack_slow HEAD | head -n 20 ··· unstacking.Unstacking.time_unstack_slow 108±3ms

Pandas asv profile unstacking.Unstacking.time_unstack_pandas_slow master | head -n 20 ··· unstacking.Unstacking.time_unstack_pandas_slow 207±10ms

Are we OK with the claim vs pandas? I think it's important that we make accurate comparisons (both good and bad) but open-minded if it seems a bit aggressive. Worth someone reviewing the code in the benchmark to ensure I haven't made a mistake.

Any final comments before merging?

If anyone here has time to review https://github.com/dask/dask/pull/7033 that would be greatly appreciated :)

As discussed in dev meeting, https://github.com/dask/dask/pull/7033 would allow dask to use the fast path, and likely eventually for our existing path to be removed

Tests now pass after merging master, not sure whether the previous tests were flaky vs something upstream changed...

Ready for a final review

This still seems to be getting a bunch of pint failures, like this one: https://dev.azure.com/xarray/xarray/_build/results?buildId=4657&view=ms.vss-test-web.build-test-results-tab&runId=73796&resultId=110407&paneView=debug

I confused, since this PR now has no mention of pint and I don't see any mention of unstacking in those test failures. I suspect I'm missing something. Any ideas for what could be happening?

Until https://github.com/pydata/xarray/pull/4751 is resolved, I've taken out the explicit pint check and replaced with a numpy check.

The code is a bit messy now — now two levels of comments. But I've put references in, so it should be tractable. Lmk any feedback.

Otherwise this is ready to go from my end.

I'm not sure whether I'm making some very basic mistake, but I'm seeing what seems like a very surprising error.

After the most recent commit, which seems to do very little apart from import pint iff it's available: https://github.com/pydata/xarray/pull/4746/commits/b33adedbfbd92df0f4188568691c7e2915bf8c19, I'm getting a lot of pint errors, unrelated to unstack / stack.

Here's the results from the run prior: https://dev.azure.com/xarray/xarray/_build/results?buildId=4650&view=ms.vss-test-web.build-test-results-tab

And from this run: https://dev.azure.com/xarray/xarray/_build/results?buildId=4651&view=ms.vss-test-web.build-test-results-tab

Any ideas what's happening? As ever, 30% chance that I made a obvious typo that I can't see... Thanks in advance.

Thank you @jthielen !

I imagine I'm making some basic error here, but what's the best approach for evaluating whether an array is a pint array?

isinstance(self.data, unit_registry.Quantity) returns False, though seems to be what we do in test_units.py?

(Pdb) unit_registry <pint.registry.UnitRegistry object at 0x13d4574c0> (Pdb) unit_registry.Quantity <class 'pint.quantity.build_quantity_class.<locals>.Quantity'> (Pdb) isinstance(self.data, unit_registry.Quantity) False (Pdb) self.data <Quantity([ 0. 0.20408163 0.40816327 0.6122449 0.81632653 1.02040816 1.2244898 1.42857143 1.63265306 1.83673469 2.04081633 2.24489796 2.44897959 2.65306122 2.85714286 3.06122449 3.26530612 3.46938776 3.67346939 3.87755102 4.08163265 4.28571429 4.48979592 4.69387755 4.89795918 5.10204082 5.30612245 5.51020408 5.71428571 5.91836735 6.12244898 6.32653061 6.53061224 6.73469388 6.93877551 7.14285714 7.34693878 7.55102041 7.75510204 7.95918367 8.16326531 8.36734694 8.57142857 8.7755102 8.97959184 9.18367347 9.3877551 9.59183673 9.79591837 10. ], 'meter')> (Pdb) self.data.__class__ <class 'pint.quantity.build_quantity_class.<locals>.Quantity'>

Thanks for responding @jthielen .

Yes, so full_like isn't creating a pint array:

(Pdb) self.data <Quantity([ 0. 0.20408163 0.40816327 0.6122449 0.81632653 1.02040816 1.2244898 1.42857143 1.63265306 1.83673469 2.04081633 2.24489796 2.44897959 2.65306122 2.85714286 3.06122449 3.26530612 3.46938776 3.67346939 3.87755102 4.08163265 4.28571429 4.48979592 4.69387755 4.89795918 5.10204082 5.30612245 5.51020408 5.71428571 5.91836735 6.12244898 6.32653061 6.53061224 6.73469388 6.93877551 7.14285714 7.34693878 7.55102041 7.75510204 7.95918367 8.16326531 8.36734694 8.57142857 8.7755102 8.97959184 9.18367347 9.3877551 9.59183673 9.79591837 10. ], 'meter')> (Pdb) np.full_like(self.data, fill_value=fill_value) array([nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan, nan]) (Pdb) fill_value nan

I do think this is a bit surprising — while fill_value isn't typed, it's compatible with the existing type.

For the moment, I'll direct pint arrays to take the existing code path — I'm more confident that we don't want to special-case pint in the unstack routine.

@keewis any thoughts on the pint issue?

Would anyone be able be familiar with this pint error? https://dev.azure.com/xarray/xarray/_build/results?buildId=4650&view=ms.vss-test-web.build-test-results-tab&runId=73654&resultId=111454&paneView=debug

It seems to be failing on the assignment: data[(..., *indexer)] = reordered, rather than anything specific to unstacking.

Here's the stack trace from there:

```python /home/vsts/work/1/s/xarray/core/variable.py:1627: in _unstack_once data[(..., *indexer)] = reordered /home/vsts/work/1/s/xarray/core/common.py:131: in array return np.asarray(self.values, dtype=dtype) /home/vsts/work/1/s/xarray/core/variable.py:543: in values return _as_array_or_item(self._data) /home/vsts/work/1/s/xarray/core/variable.py:275: in _as_array_or_item data = np.asarray(data) /usr/share/miniconda/envs/xarray-tests/lib/python3.8/site-packages/numpy/core/_asarray.py:83: in asarray return array(a, dtype, copy=False, order=order)

self = <Quantity([ 0 0 0 0 0 1 1 1 1 1 2 2 2 2 2 3 3 3 3 3 4 4 4 4 4 5 5 5 5 5 6 6 6 6 6 7 7 7 7 7 8 8 8 8 8 9 9 9 9 10], 'meter')> t = None

def __array__(self, t=None):

  warnings.warn(
        "The unit of the quantity is stripped when downcasting to ndarray.",
        UnitStrippedWarning,
        stacklevel=2,
    )

E pint.errors.UnitStrippedWarning: The unit of the quantity is stripped when downcasting to ndarray.

/usr/share/miniconda/envs/xarray-tests/lib/python3.8/site-packages/pint/quantity.py:1683: UnitStrippedWarning ```

Worst case, I can direct pint arrays to the existing unstack path, but ideally this would work.

Great, thanks, I'm making that change.

Is there any need to keep the sparse kwarg? My inclination is to remove it and retain types — so to get a sparse array back, convert to sparse before unstacking?

Very nice!

I think we solve the issue with sparse by using np.full_like, which does dispatching on NumPy 1.17+ via NEP-18 __array_function__ (which I'm pretty sure sparse supports).

The bigger challenge that I'm concerned about here are dask arrays, which don't support array assignment like this at all (https://github.com/dask/dask/issues/2000). We will probably need to keep around the slower option for dask, at least for now.

Any ideas on how sparse arrays should be handled in unstack? Currently we use reindex, so this seems to pass through without much effort on our part.

In the new code, we're creating an array with np.full and then assigning to the appropriate locations. Can we do something similar that's not dependent on the underlying numpy / sparse / dask storage?