issue_comments

I'll drop a PR, it might be easier to try and play with this than a piece of code lost in an issue.

Okay I could re do my test. If I manually call compute() before doing isel(......) my whole computation takes about 5.65 seconds. However if I try with my naive patch it takes 32.34 seconds.

I'm sorry I cannot share as is my code, the relevant portion is really in the middle of many things. I'll try to get a minimalist version of it to share with you.

@pl-marasco Ok that's strange. I should have saved my use case :/ I will try to reproduce it and will provide a gist of it soon.

@pl-marasco Thanks for the example ! With it I have the same result as you, it takes the same time with patch or with compute.

However, I could construct an example giving very different results. It is quite close to my original code:

``` time_start = time.perf_counter() COORDS = dict( time=pd.date_range("2042-01-01", periods=200, freq=pd.DateOffset(days=1)), ) da = xr.DataArray( np.random.rand(200 * 3500 * 350).reshape((200, 3500, 350)), dims=('time', 'x', 'y'), coords=COORDS ).chunk(dict(time=-1, x=100, y=100))

resampled = da.resample(time="MS")

for label, sample in resampled:
    # sample = sample.compute()
    idx = sample.argmax('time')
    sample.isel(time=idx)

time_elapsed = time.perf_counter() - time_start
print(time_elapsed, " secs")

``` (Basically I want for each month the first event occurring in it).

Without the patch and uncommenting sample = sample.compute(), it takes 5.7 secs. With the patch it takes 53.9 seconds.

I wrote a very naive fix, it works but seems to perform really slowly, I would appreciate some feedback (I'm a beginner with Dask). Basically, I added k = dask.array.asarray(k, dtype=np.int64) to do the exact same thing as with numpy. I can create a PR if it's better to review this

The patch: ``` class VectorizedIndexer(ExplicitIndexer): """Tuple for vectorized indexing.

All elements should be slice or N-dimensional np.ndarray objects with an
integer dtype and the same number of dimensions. Indexing follows proposed
rules for np.ndarray.vindex, which matches NumPy's advanced indexing rules
(including broadcasting) except sliced axes are always moved to the end:
https://github.com/numpy/numpy/pull/6256
"""

__slots__ = ()

def __init__(self, key):
    if not isinstance(key, tuple):
        raise TypeError(f"key must be a tuple: {key!r}")

    new_key = []
    ndim = None
    for k in key:
        if isinstance(k, slice):
            k = as_integer_slice(k)
        elif isinstance(k, np.ndarray) or isinstance(k, dask.array.Array):
            if not np.issubdtype(k.dtype, np.integer):
                raise TypeError(
                    f"invalid indexer array, does not have integer dtype: {k!r}"
                )
            if ndim is None:
                ndim = k.ndim
            elif ndim != k.ndim:
                ndims = [k.ndim for k in key if isinstance(k, np.ndarray)]
                raise ValueError(
                    "invalid indexer key: ndarray arguments "
                    f"have different numbers of dimensions: {ndims}"
                )
            if isinstance(k, dask.array.Array):
                k = dask.array.asarray(k, dtype=np.int64)
            else:
                k = np.asarray(k, dtype=np.int64)
        else:
            raise TypeError(
                f"unexpected indexer type for {type(self).__name__}: {k!r}"
            )
        new_key.append(k)

    super().__init__(new_key)

```