html_url,issue_url,id,node_id,user,created_at,updated_at,author_association,body,reactions,performed_via_github_app,issue
https://github.com/pydata/xarray/issues/2459#issuecomment-426646669,https://api.github.com/repos/pydata/xarray/issues/2459,426646669,MDEyOklzc3VlQ29tbWVudDQyNjY0NjY2OQ==,5635139,2018-10-03T13:55:40Z,2018-10-03T16:13:41Z,MEMBER,"My working hypothesis is that pandas has a set of fast routines in C, such that it can stack without reindexing to the full index. The routines only work in 1-2 dimensions.

So without some hackery (i.e. converting multi-dimensional arrays to pandas' size and back), the current implementation is reasonable*. Next step would be to write our own routines that can operate on multiple dimensions (numbagg!).

Is that consistent with others' views, particularly those who know this area well?

'* one small fix that would improve performance of `series.to_xarray()` only, is the [comment above](https://github.com/pydata/xarray/issues/2459#issuecomment-426483497). Lmk if you think worth making that change","{""total_count"": 0, ""+1"": 0, ""-1"": 0, ""laugh"": 0, ""hooray"": 0, ""confused"": 0, ""heart"": 0, ""rocket"": 0, ""eyes"": 0}",,365973662
https://github.com/pydata/xarray/issues/2459#issuecomment-426483497,https://api.github.com/repos/pydata/xarray/issues/2459,426483497,MDEyOklzc3VlQ29tbWVudDQyNjQ4MzQ5Nw==,5635139,2018-10-03T01:30:07Z,2018-10-03T01:30:07Z,MEMBER,"It's 3x faster to unstack & stack all-but-one level, vs reindexing over a filled-out index (and I think always produces the same result). 

Our current code takes the slow path.

I could make that change, but that strongly feels like I don't understand the root cause. I haven't spent much time with reshaping code - lmk if anyone has ideas.

```python

idx = cropped.index
full_idx = pd.MultiIndex.from_product(idx.levels, names=idx.names)

reindexed = cropped.reindex(full_idx)

%timeit reindexed = cropped.reindex(full_idx)
# 1 loop, best of 3: 278 ms per loop

%%timeit
stack_unstack = (
    cropped
    .unstack(list('yz'))
    .stack(list('yz'),dropna=False)
)
# 10 loops, best of 3: 80.8 ms per loop

stack_unstack.equals(reindexed)
# True
```","{""total_count"": 0, ""+1"": 0, ""-1"": 0, ""laugh"": 0, ""hooray"": 0, ""confused"": 0, ""heart"": 0, ""rocket"": 0, ""eyes"": 0}",,365973662
https://github.com/pydata/xarray/issues/2459#issuecomment-426408924,https://api.github.com/repos/pydata/xarray/issues/2459,426408924,MDEyOklzc3VlQ29tbWVudDQyNjQwODkyNA==,5635139,2018-10-02T19:57:20Z,2018-10-02T19:57:20Z,MEMBER,"When I stepped through, it was by-and-large all taken up by https://github.com/pydata/xarray/blob/master/xarray/core/dataset.py#L3121. That's where the boxing & unboxing of the datetimes is from.

I haven't yet discovered how the alternative path avoids this work. If anyone has priors please lmk!","{""total_count"": 0, ""+1"": 0, ""-1"": 0, ""laugh"": 0, ""hooray"": 0, ""confused"": 0, ""heart"": 0, ""rocket"": 0, ""eyes"": 0}",,365973662