xarray
xarray copied to clipboard
pydata
[Bug]: reading NaT/NaN on M1 ARM chip
What happened?
I have nan values in a date vector stored in a netCDF. When I read on my ARM Apple computer with xr.open_dataset(), it is not properly recognized.
For example, the following data is stored in a NetCDF:
date = pd.date_range(...)
date[4] = nan
Then when I read the file:
date[4] is set to date[0], which is the first date of the range instead of a 'NaT'.
I understand that this issue is quite weird and it doesn't seem to happen on other OS. Actually, I try on MacOS (with an intel processor) and on two different Linux computers, and in those configurations, date[4] is properly set to 'NaT' after opening the netCDF with xr.open_dataset(). Note that I tried with the same version of xarray as well as with different versions, and I just can't seem to reproduce this issue on any machine except on the M1 ARM chip.
What did you expect to happen?
I expect the following result after running the minimal example:
array(['2022-01-01T00:00:00.000000000', '2022-01-02T00:00:00.000000000',
'2022-01-03T00:00:00.000000000', '2022-01-04T00:00:00.000000000',
'NaT', '2022-01-06T00:00:00.000000000',
'2022-01-07T00:00:00.000000000', '2022-01-08T00:00:00.000000000',
'2022-01-09T00:00:00.000000000', '2022-01-10T00:00:00.000000000'],
dtype='datetime64[ns]')
Minimal Complete Verifiable Example
import xarray as xr
import pandas as pd
import numpy as np
time = pd.date_range(start="2022-01-01",end="2022-01-10").to_pydatetime()
time[4] = np.datetime64("NaT")
ds = xr.Dataset(
data_vars=dict(
time=(["nt"], time),
),
)
ds.to_netcdf('test.nc')
ds_r = xr.open_dataset('test.nc')
ds_r.time
Relevant log output
array(['2022-01-01T00:00:00.000000000', '2022-01-02T00:00:00.000000000',
'2022-01-03T00:00:00.000000000', '2022-01-04T00:00:00.000000000',
'2022-01-01T00:00:00.000000000', '2022-01-06T00:00:00.000000000',
'2022-01-07T00:00:00.000000000', '2022-01-08T00:00:00.000000000',
'2022-01-09T00:00:00.000000000', '2022-01-10T00:00:00.000000000'],
dtype='datetime64[ns]')
Anything else we need to know?
No response
Environment
INSTALLED VERSIONS
commit: None python: 3.10.1 | packaged by conda-forge | (main, Dec 22 2021, 01:38:36) [Clang 11.1.0 ] python-bits: 64 OS: Darwin OS-release: 21.2.0 machine: arm64 processor: arm byteorder: little LC_ALL: None LANG: en_US.UTF-8 LOCALE: ('en_US', 'UTF-8') libhdf5: 1.12.1 libnetcdf: 4.8.1
xarray: 0.20.2 pandas: 1.3.5 numpy: 1.21.5 scipy: 1.7.3 netCDF4: 1.5.8 pydap: None h5netcdf: None h5py: None Nio: None zarr: None cftime: 1.5.1.1 nc_time_axis: None PseudoNetCDF: None rasterio: None cfgrib: None iris: None bottleneck: None dask: 2021.12.0 distributed: 2021.12.0 matplotlib: 3.5.1 cartopy: 0.20.1 seaborn: None numbagg: None fsspec: 2021.11.1 cupy: None pint: None sparse: None setuptools: 60.0.4 pip: 21.3.1 conda: None pytest: None IPython: 8.0.0 sphinx: None
Thanks @philippemiron .
My guess is that this is an issue with an underlying library, since xarray doesn't generally do these operations in its code. Do you know if there are any similar issues in libnetcdf or netCDF4?
(Others know more than me about these libraries, so please feel free to interject)
I tried reproducing on an M1 Mac, but my install of python seems to report that it's on an x86_64 (version='Darwin Kernel Version 21.0.1: Tue Sep 14 20:56:24 PDT 2021 ; root:xnu-8019.30.61~4/RELEASE_ARM64_T6000', machine='x86_64'). It didn't reproduce, unsurprisingly.
Does uninstalling netCDF4 help? That would isolate it to that library and its dependencies.
I'm actually using miniforge which natively supports ARM64. Uninstalling netCDF4 does not fix the issue. And actually, opening the same file as follow:
from netCDF4 import Dataset
f = Dataset('test.nc')
f['time'][:]
gives the expected results (dates are not recognized but the nan is there):
masked_array(data=[0.0, 1.0, 2.0, 3.0, --, 5.0, 6.0, 7.0, 8.0, 9.0],
mask=[False, False, False, False, True, False, False, False,
False, False],
fill_value=nan)
I sorted out my M1 python installation and can reproduce:
In [21]: ds_r.time
Out[21]:
<xarray.DataArray 'time' (nt: 10)>
array(['2022-01-01T00:00:00.000000000', '2022-01-02T00:00:00.000000000',
'2022-01-03T00:00:00.000000000', '2022-01-04T00:00:00.000000000',
'2022-01-01T00:00:00.000000000', '2022-01-06T00:00:00.000000000', # Note the first value on this line!
'2022-01-07T00:00:00.000000000', '2022-01-08T00:00:00.000000000',
'2022-01-09T00:00:00.000000000', '2022-01-10T00:00:00.000000000'],
dtype='datetime64[ns]')
Dimensions without coordinates: nt
It's quite surprising we get '2022-01-01T00:00:00.000000000' rather than NaT — why the beginning of the year?!
I suspect it's not directly an xarray issue given Xarray is only python code, and python code does not directly branch by CPU. I've frequently had issues like this where it's difficult to understand which library is responsible, I'd welcome any more investigation here.
It is replaced by the first value of the array. If you change to:
time = pd.date_range(start="2022-01-02",end="2022-01-11").to_pydatetime()
the NaT is replaced by '2022-01-02T00:00:00.000000000'. Maybe it is stored internally as a time_origin and some time_delta, and the NaT are replaced by 0?
I got caught by this one yesterday on an M1 machine. I did some digging and found what I think to be the underlying issue. The short explanation is that the time conversion functions do an astype(np.int64) or equivalent cast on arrays that contain nans. This is undefined behavior and very soon, doing this will start to emit RuntimeWarnings.
I knew from my own data files that it wasn't the first element of the array being substituted but whatever was in the units as the epoch. I started to poke at the xarray internals (and the CFtime internals) to try to get a minimal example working, eventually found the following:
On an M1:
>>> from xarray.coding.times import _decode_datetime_with_pandas
>>> import numpy as np
>>> _decode_datetime_with_pandas(np.array([20000, float('nan')]), "days since 1950-01-01", "proleptic_gregorian")
array(['2004-10-04T00:00:00.000000000', '1950-01-01T00:00:00.000000000'],
dtype='datetime64[ns]')
>>> np.array(np.nan).astype(np.int64)
array(0)
On an x86_64:
>>> from xarray.coding.times import _decode_datetime_with_pandas
>>> import numpy as np
>>> _decode_datetime_with_pandas(np.array([20000, float('nan')]), "days since 1950-01-01", "proleptic_gregorian")
array(['2004-10-04T00:00:00.000000000', 'NaT'],
dtype='datetime64[ns]')
>>> np.array(np.nan).astype(np.int64)
array(-9223372036854775808)
This issue is not Apple/M1/clang specific, I tested on an aws graviton (arm) instance and got the same results with ubuntu/gcc:
Python 3.10.4 (main, Jun 29 2022, 12:14:53) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from xarray.coding.times import _decode_datetime_with_pandas
>>> import numpy as np
>>> _decode_datetime_with_pandas(np.array([20000, float('nan')]), "days since 1950-01-01", "proleptic_gregorian")
array(['2004-10-04T00:00:00.000000000', '1950-01-01T00:00:00.000000000'],
dtype='datetime64[ns]')
>>> np.array(np.nan).astype(np.int64)
array(0)
Here is where the cast is happening on the internal xarray implementation, CFtime has similar casts in its implementation. https://github.com/pydata/xarray/blob/8417f495e6b81a60833f86a978e5a8080a619aa0/xarray/coding/times.py#L237-L239
Some additional info for when how to figure out the best way to address this.
For the decode using pandas approach, two things I tried worked: using a pandas.array with a nullable integer data type, or simulating what happens on x86_64 systems by checking for nans in the incoming array and setting those positions to numpy.iinfo(np.int64).min.
the pandas nullable integer array:
# note that is a capital i Int64 to use the nullable type.
flat_num_dates_ns_int = pd.array(flat_num_dates * _NS_PER_TIME_DELTA[delta], dtype="Int64")
simulate x86:
flat_num_dates_ns_int = (flat_num_dates * _NS_PER_TIME_DELTA[delta]).astype(
np.int64
)
flat_num_dates_ns_int[np.isnan(flat_num_dates)] = np.iinfo(np.int64).min
The pandas solution is explicitly experimental in their docs, and the emulate version just feels "hacky" to me. These don't break any existing tests on my local machine.
cftime itself has no support for nan type missing values and will fail:
(on x86_64)
>>> import numpy as np
>>> from xarray.coding.times import decode_cf_datetime
>>> decode_cf_datetime(np.array([0, np.nan]), "days since 1950-01-01", use_cftime=True)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/abarna/.pyenv/versions/3.8.5/lib/python3.8/site-packages/xarray/coding/times.py", line 248, in decode_cf_datetime
dates = _decode_datetime_with_cftime(flat_num_dates, units, calendar)
File "/home/abarna/.pyenv/versions/3.8.5/lib/python3.8/site-packages/xarray/coding/times.py", line 164, in _decode_datetime_with_cftime
cftime.num2date(num_dates, units, calendar, only_use_cftime_datetimes=True)
File "src/cftime/_cftime.pyx", line 484, in cftime._cftime.num2date
TypeError: unsupported operand type(s) for +: 'cftime._cftime.DatetimeGregorian' and 'NoneType'
cftime is happy with masked arrays:
>>> import cftime
>>> a1 = np.ma.masked_invalid(np.array([0, np.nan]))
>>> cftime.num2date(a1, "days since 1950-01-01")
masked_array(data=[cftime.DatetimeGregorian(1950, 1, 1, 0, 0, 0, 0), --],
mask=[False, True],
fill_value='?',
dtype=object)