issues: 419945711

Code Sample

Unfortunately I cannot include the original data, as it is quite large, but I can make an accessible dropbox folder if needed.

```python debug = False if not debug: import matplotlib matplotlib.use('Agg')

import matplotlib.pyplot as plt import xarray as xr import metpy.calc as mpcalc from metpy.units import units from glob import glob import numpy as np import pandas as pd from multiprocessing import Pool from functools import partial import os from utils import * import sys

variable_name = 'cape_cin' projections = ['de','it','nord']

def main(): """In the main function we basically read the files and prepare the variables to be plotted. This is not included in utils.py as it can change from case to case.""" files = glob(input_file) dset = xr.open_mfdataset(files) # Only take hourly data dset = dset.sel(time=pd.date_range(dset.time[0].values, dset.time[-1].values, freq='H')) # Parse metadata for metpy dset = dset.metpy.parse_cf() # Select variable cape = dset['CAPE_ML'].squeeze().load()

# Get coordinates
lon, lat = get_coordinates(dset)
lon2d, lat2d = np.meshgrid(lon, lat)

time = pd.to_datetime(dset.time.values)
cum_hour=np.array((time-time[0]) / pd.Timedelta('1 hour')).astype("int")

levels_cape = np.arange(250., 2000., 250.)

cmap = truncate_colormap(plt.get_cmap('gist_stern_r'), 0., 0.7)

for projection in projections:
    fig = plt.figure(figsize=(figsize_x, figsize_y))
    ax  = plt.gca()        
    m, x, y =get_projection(lon2d, lat2d, projection, labels=True)

    # All the arguments that need to be passed to the plotting function
    args=dict(m=m, x=x, y=y, ax=ax, cmap=cmap,
             cape=cape, levels_cape=levels_cape,
             time=time, projection=projection, cum_hour=cum_hour)

    print('Pre-processing finished, launching plotting scripts')
    if debug:
        plot_files(time[1:2], **args)
    else:
        # Parallelize the plotting by dividing into chunks and processes 
        dates = chunks(time, chunks_size)
        plot_files_param=partial(plot_files, **args)
        p = Pool(processes)
        p.map(plot_files_param, dates)

def plot_files(dates, **args): first = True for date in dates: # Find index in the original array to subset when plotting i = np.argmin(np.abs(date - args['time'])) # Build the name of the output image filename = subfolder_images[args['projection']]+'/'+variable_name+'_%s.png' % args['cum_hour']

    # Do the plot 
    cs = args['ax'].contourf(args['x'], args['y'], args['cape'][i], 
                                             extend='both', cmap=args['cmap'],
                                             levels=args['levels_cape'])

    if first:
        plt.colorbar(cs, orientation='horizontal', label='CAPE [J/kg]', pad=0.03, fraction=0.04)

    if debug:
        plt.show(block=True)
    else:
        plt.savefig(filename, **options_savefig)

    remove_collections([cs, an_fc, an_var, an_run])

    first = False

if name == "main": main() ```

Problem description

My task is to routinely plot output from weather models in an automatized way. The main driver script is written in 'bash' and calls different 'python' scripts at the same time.

Given that the input data is always dimensioned '(time, level, latitude, longitude)' and that I have to do one plot per timestep it seemed natural to me to use multiprocessing to split the load of different time-chunks over different processors. Since I also have to pass some parameters/variables I call the function to do the actual plotting plot_files using the partial constructor plot_files_param=partial(plot_files, **args). This way I can then unpack all the needed parameters/variables in the function itself without needing to declare them as global variables. Note that this is fundamental as, to make everything fast, I only create most of the variables/instances once and then update them at every iteration.

This works fine if I load the input netcdf file with open_dataset, but instead if I use open_mfdataset (as in some case I have many files scattered in a folder) then I have to be careful in explicitly loading the arrays before passing them to the plotting function, otherwise the script will hang.

More specifically. if I remove the .load() statement from this line python cape = dset['CAPE_ML'].squeeze().load() then cape stays a dask array and gets passed as argument to plot_files. Then I was expecting the plot call to load it as array in memory before doing the plot, but this does not happen (I think) and then the script just hangs (does not post any error). Interestingly enough, if I use the option debug = True, which instead of parallelizing the plotting just plot 1 timestep and directly show the output, everything works fine, although the array passed is still a dask array. So, this is definitely something related to the use of the partial construct.

In theory I would like to keep the .load() usage as small as possible as data can be pretty huge and I only want to load it into memory when I have to plot it.

It may well be that my code is somehow malformed but as far as I know this should not be the expected behaviour. Please apologize if I'm doing something wrong; every suggestion is highly appreciated! Thought that I report the error as I found it to be non-expected behaviour.

Expected Output

Array should load into memory when plotting without the need to do it explicitly.

Output of xr.show_versions()

INSTALLED VERSIONS ------------------ commit: None python: 2.7.15 | packaged by conda-forge | (default, Nov 29 2018, 06:43:57) [GCC 4.8.2 20140120 (Red Hat 4.8.2-15)] python-bits: 64 OS: Linux OS-release: 3.16.0-7-amd64 machine: x86_64 processor: byteorder: little LC_ALL: en_US.UTF-8 LANG: en_US.UTF-8 LOCALE: None.None libhdf5: 1.10.3 libnetcdf: 4.6.1 xarray: 0.11.2 pandas: 0.23.4 numpy: 1.15.4 scipy: 1.2.0 netCDF4: 1.4.2 pydap: None h5netcdf: None h5py: 2.8.0 Nio: None zarr: None cftime: 1.0.2.1 PseudonetCDF: None rasterio: None cfgrib: 0.9.5.1 iris: None bottleneck: None cyordereddict: None dask: 1.0.0 distributed: 1.25.0 matplotlib: 2.2.3 cartopy: 0.17.0 seaborn: 0.9.0 setuptools: 40.6.2 pip: 18.1 conda: None pytest: None IPython: 5.8.0 sphinx: None