Negative timesteps after .to_netcdf with long time periods?

[cwerner]

Hi.

I discovered that I get negative times on the time dimension for long intervals (years 1700-2013, monthly timestep).

import xray
import numpy as np
import pandas as pd

years = range(1700,2014)
LATS = np.arange(-89.75, 90.0, 0.5)
LONS = np.arange(-179.75, 180.0, 0.5)

tlist = pd.date_range('%d-01-01' % years[0], periods=12*len(years), freq='M')

da = xray.DataArray(np.ones((12*len(years), 360, 720))*-9999, \
            [('time', tlist), ('latitude',  LATS), ('longitude', LONS) ])

# i then fill the dataarray with info from a text file (using read_csv from pandas)
# eventually I dump to netcdf

ds = xray.Dataset({"mgpp": da})
ds.to_netcdf('test_%d-%d.nc' % (years[0], years[-1]))

If I "ncdump -c mgpp_1700-2013.nc I get:

netcdf mgpp_1700-2013 {
dimensions:
    latitude = 360 ;
    time = 3768 ;
    longitude = 720 ;
variables:
    float latitude(latitude) ;
    float mgpp(time, latitude, longitude) ;
        mgpp:units = "gCm-2" ;
    float longitude(longitude) ;
    float time(time) ;
        time:units = "days since 1700-01-31 00:00:00" ;
        time:calendar = "proleptic_gregorian" ;
data:

time = 0, 28, 59, 89, 120, 150, 181, 212, 242, 273, 303, 334, 365, 393, 424, 
    454, 485, 515, 546, 577, 607, 638, 668, 699, 730, 758, 789, 819, 850, 
    880, 911, 942, 972, 1003, 1033, 1064, 1095, 1123, 1154, 1184, 1215, 1245, 
    1276, 1307, 1337, 1368, 1398, 1429, 1460, 1489, 1520, 1550, 1581, 1611, 
    1642, 1673, 1703, 1734, 1764, 1795, 1826, 1854, 1885, 1915, 1946, 1976, 
    2007, 2038, 2068, 2099, 2129, 2160, 2191, 2219, 2250, 2280, 2311, 2341, 
    2372, 2403, 2433, 2464, 2494, 2525, 2556, 2584, 2615, 2645, 2676, 2706, 
    2737, 2768, 2798, 2829, 2859, 2890, 2921, 2950, 2981, 3011, 3042, 3072, 
    3103, 3134, 3164, 3195, 3225, 3256, 3287, 3315, 3346, 3376, 3407, 3437, 
    3468, 3499, 3529, 3560, 3590, 3621, 3652, 3680, 3711, 3741, 3772, 3802, (...)

and eventually:

  (...) 106435, 106466, 106497, 106527, 106558, 106588, 106619, 106650, 106679, 
    106710, 106740, -106732.982337963, -106702.982337963, -106671.982337963, 
    -106640.982337963, -106610.982337963, -106579.982337963, 
    -106549.982337963, -106518.982337963, -106487.982337963, 
    -106459.982337963, -106428.982337963, -106398.982337963, 
    -106367.982337963, -106337.982337963, -106306.982337963, 
    -106275.982337963, -106245.982337963, -106214.982337963, 
    -106184.982337963, -106153.982337963, -106122.982337963, 
    -106094.982337963, -106063.982337963, -106033.982337963, 
    -106002.982337963, -105972.982337963, -105941.982337963, 
    -105910.982337963, -105880.982337963, -105849.982337963, (...)

Not sure if I can inflence that at "dump" time with to_netcdf? I know about the time limitation, but my years should be non-critical, no?

[shoyer]

This looks like some sort of integer overflow bug. Oddly, the times actually roundtrips fine back into xray (but probably not other programs that read netCDFs).

[stale]

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

[spencerkclark]

This is the same issue as #2272, which was fixed in #2519.