In order to maintain a list of currently relevant issues, we mark issues as stale after a period of inactivity If this issue remains relevant, please comment here; otherwise it will be marked as closed automatically

For reference, it may be helpful to try your example on a smaller dataset: ds = xr.Dataset() ds['data1'] = xr.DataArray(np.arange(100), coords={'t1': np.linspace(0, 1, 100)}) ds['data1b'] = xr.DataArray(np.arange(100, 200), coords={'t1': np.linspace(0, 1, 100)}) ds['data2'] = xr.DataArray(np.arange(200, 500), coords={'t2': np.linspace(0, 1, 300)}) ds['data2b'] = xr.DataArray(np.arange(600, 900), coords={'t2': np.linspace(0, 1, 300)}) This is what the data looks like: ```

ds <xarray.Dataset> Dimensions: (t1: 100, t2: 300) Coordinates: * t1 (t1) float64 0.0 0.0101 0.0202 0.0303 0.0404 0.05051 0.06061 ... * t2 (t2) float64 0.0 0.003344 0.006689 0.01003 0.01338 0.01672 ... Data variables: data1 (t1) int64 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 ... data1b (t1) int64 100 101 102 103 104 105 106 107 108 109 110 111 112 ... data2 (t2) int64 200 201 202 203 204 205 206 207 208 209 210 211 212 ... data2b (t2) int64 600 601 602 603 604 605 606 607 608 609 610 611 612 ... and here's what happens with `where`: ds.where(ds.data1 < 50, drop=True) <xarray.Dataset> Dimensions: (t1: 50, t2: 300) Coordinates: * t1 (t1) float64 0.0 0.0101 0.0202 0.0303 0.0404 0.05051 0.06061 ... * t2 (t2) float64 0.0 0.003344 0.006689 0.01003 0.01338 0.01672 ... Data variables: data1 (t1) float64 0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0 11.0 ... data1b (t1) float64 100.0 101.0 102.0 103.0 104.0 105.0 106.0 107.0 ... data2 (t2, t1) float64 200.0 200.0 200.0 200.0 200.0 200.0 200.0 ... data2b (t2, t1) float64 600.0 600.0 600.0 600.0 600.0 600.0 600.0 ... ```

I suspect this probably isn't really doing what you want, unless you really want two-dimensional versions of data2 and data2b. It probably makes more sense to subset out the relevant variables first.

Broadcasting producing gigantic arrays without any warning is really a NumPy issue, e.g., trynp.where(True, np.zeros((1000000, 1)), np.ones((1, 1000000))) (this will probably crash your computer!).

The problem isn't as bad with a smaller example (though the runtime is doubled). I've attached a minimum working example, which seems to suggest that maybe there was a problem with xarray creating a MultiIndex and duplicating all the data? (I've left in input() to allow checking memory usage before the program exists, but there isn't much difference in this example). xrmin.py.txt

Edit by @shoyer: added code from attachment inline: ```python

!/usr/bin/env python3

import time import sys import numpy as np import xarray as xr

ds = xr.Dataset() ds['data1'] = xr.DataArray(np.arange(1000), coords={'t1': np.linspace(0, 1, 1000)}) ds['data1b'] = xr.DataArray(np.arange(1000, 2000), coords={'t1': np.linspace(0, 1, 1000)}) ds['data2'] = xr.DataArray(np.arange(2000, 5000), coords={'t2': np.linspace(0, 1, 3000)}) ds['data2b'] = xr.DataArray(np.arange(6000, 9000), coords={'t2': np.linspace(0, 1, 3000)}) if sys.argv[1] == "nodrop": now = time.time() print(ds.where(ds.data1 < 50, drop=True)) print("Took {} seconds".format(time.time() - now)) elif sys.argv[1] == "drop": ds1 = ds.drop('t2') now = time.time() print(ds1.where(ds1.data1 < 50, drop=True)) print("Took {} seconds".format(time.time() - now)) input("Press return to exit") ```

. So in my case extracting the data with the shared dimension using ds.drop() is appropriate. It would be nice to have xarray throw a warning or error to prevent me chomping up all the RAM in my system if I do try to do this sort of thing though.

I'll let @shoyer give a definitive answer here, but I don't think that .where is meant to check whether the input makes "sense" or not. What happens is related to how xarray chooses to broadcast non matching dimensions:

```python import xarray as xr import numpy as np d1 = xr.DataArray(np.arange(3), coords={'t1': np.linspace(0, 1, 3)}, dims='t1') d2 = xr.DataArray(np.arange(4), coords={'t2': np.linspace(0, 1, 4)}, dims='t2')

d2 * d1 <xarray.DataArray (t2: 4, t1: 3)> array([[0, 0, 0], [0, 1, 2], [0, 2, 4], [0, 3, 6]]) Coordinates: * t2 (t2) float64 0.0 0.3333 0.6667 1.0 * t1 (t1) float64 0.0 0.5 1.0

d2.where(d1 == 1) <xarray.DataArray (t2: 4, t1: 3)> array([[ nan, 0., nan], [ nan, 1., nan], [ nan, 2., nan], [ nan, 3., nan]]) Coordinates: * t2 (t2) float64 0.0 0.3333 0.6667 1.0 * t1 (t1) float64 0.0 0.5 1.0 ``` which "makes sense", but is going to have a huge memory consumption if your arrays are large.

data1 and data2 represent two stages of data acquisition within one "shot" of our experiment. I'd like to be able to group each shot's data into a single dataset.

I want to extract from the dataset only the values for which my where() condition is true, and I'll only be using DataArrays which share the same dimension as the one in the condition. For example, if I do: ds_low = ds.where(ds.data1 < 0.1, drop=True) I'll only use stuff in ds_low with the same dimension as ds.data1. So in my case extracting the data with the shared dimension using ds.drop(<unused dim>) is appropriate.

It would be nice to have xarray throw a warning or error to prevent me chomping up all the RAM in my system if I do try to do this sort of thing though. Or it could simply mask off with NaN everything in the DataArrays which have a different dimension.

Give me a second to provide a minimal working example.

Thanks for the report! It would be great if you could be a bit more specific: - if data1 and data2 are unrelated, why do you want to apply where on both variables? What is your expectation on the output? - do you have the possibility to produce a minimal, self-contained working example?