Frequently Asked Questions

wrf.getvar() uses more memory when xarray is enabled. Is there a memory leak?

The difference in memory usage can be observed when running the following code with and without xarray enabled (see wrf.disable_xarray()):

from netCDF4 import Dataset
import wrf

# Uncomment to disable xarray
# wrf.disable_xarray()

for i in range(150):
    f = Dataset('wrfout_d01_2005-07-17_12_00_00.nc')
    p = wrf.getvar(f, 'pressure')
    f.close()

When xarray is enabled, there is an internal thread-local cache used to hold the XLAT and XLONG coordinate variables, along with a boolean variable to remember if a particular file or sequence is from a moving nest. This is useful when working with sequences of WRF files, so that the XLAT and XLONG variables aren’t repeatedly extracted. This cache is limited to holding 20 items by default, where the key to each cache entry is the object ID for the file or sequence of files. In this particular example, a new file object is created for each iteration, which will have a new object ID, so the 20 cache spots are going to be filled up. Memory usage will be quite a bit higher for the xarray-enabled case, but the memory usage will eventually top off.

There is a function wrf.set_cache_size() that can be used to decrease the number of files/sequences in the cache. By setting the cache size to 0, the cache can be disabled completely.

Can I use xarray.Dataset as an input to the wrf-python functions?

Currently, wrf-python is designed to use the API for PyNIO and netCDF4-python, which differs from the xarray.Dataset API. However, there is an undocumented feature for Dataset objects that can be used as workaround. Each Dataset object has a xarray.Dataset._file_obj.ds attribute that can be used to obtain the underlying netCDF file object. Keep in mind that this is not in the public API for xarray, so it could break in a future release.

Here is an example:

import xarray
import wrf

ds = xarray.open_dataset(FILENAME)
slp = wrf.getvar(ds._file_obj.ds, "slp")

In a future release of wrf-python, direct support for Dataset objects will be added and this will no longer be necessary.