issue_comments

Hi @benbovy,

Thanks for the detailed response.

Yeah, that it was only raising for the second multi indexing map, does seem like a bug in that case, I'll leave the ticket open to track that.

I didn't stumble on the set_levels function while skimming though the docs, thanks. I've updated my function to make use of that. I'm hoping things should be safe with this, and I'm correctly replacing things in the correct order.

For anyone else who's looking to do the same, or for anyone to tell me what I'm doing is not safe, or there's a simpler way, here's the updated function:

```python import numpy as np import pandas as pd import xarray as xr

def map_coords(ds, *, name, mapping): """ Takes a xarray dataset's coordinate values and updates them with the given the provided mapping. In-place.

Can handle both regular indices and multi-level indices.

ds: An xr.Datset
name: name of the coordinate to update
mapping: dictionary, key of old value, value of new value.
"""
# all attrs seem to get dropped on coords even if only
# one is altered. Hold on to them and reapply after
coord_attrs = {c: ds[c].attrs for c in ds.coords}

if ds.indexes.is_multi(name):
    target = name
    parent = ds.indexes[target].name

    if target == parent:
        valid_targets = ds.indexes[parent].names
        raise ValueError(
            f"Can only map levels of a MultiIndex, not the MultiIndex "
            f"itself. Target one of {valid_targets}",
        )

    multi_index = ds.indexes[parent]
    level_values = dict(zip(multi_index.names, multi_index.levels))
    new_values = [mapping[v] for v in level_values[name]]
    ds.coords[parent] = multi_index.set_levels(new_values, level=target)
else:
    old_values = ds.coords[name].values
    new_values = [mapping[v] for v in old_values]
    ds[name] = new_values

# reapply attrs
for coord, attrs in coord_attrs.items():
    ds[coord].attrs = attrs

midx = pd.MultiIndex.from_product([list("abc"), [0, 1]], names=("x_one", "x_two")) midy = pd.MultiIndex.from_product([list("abc"), [0, 1]], names=("y_one", "y_two")) mda = xr.DataArray(np.random.rand(6, 6, 3), [("x", midx), ("y", midy), ("z", range(3))])

map_coords(mda, name="z", mapping={0: "zero", 1: "one", 2: "two"}) map_coords(mda, name="x_one", mapping={"a": "aa", "b": "bb", "c": "cc"}) map_coords(mda, name="y_one", mapping={"a": "aa", "b": "bb", "c": "cc"})

print(mda) ```

@rabernat Still seem to get a SIGKILL 9 (exit code 137) when trying to run with that pre-processor as well.

Maybe my expectations of how it lazy loads files is too high. The machine I'm running on has 8GB or ram and the files in total are just under 1Tb

@jhamman Here's the ncdump of one of the resource files:

```bash netcdf \34.128_1900_01_05_05 { dimensions: longitude = 720 ; latitude = 361 ; time = UNLIMITED ; // (124 currently) variables: float longitude(longitude) ; longitude:units = "degrees_east" ; longitude:long_name = "longitude" ; float latitude(latitude) ; latitude:units = "degrees_north" ; latitude:long_name = "latitude" ; int time(time) ; time:units = "hours since 1900-01-01 00:00:0.0" ; time:long_name = "time" ; time:calendar = "gregorian" ; short sst(time, latitude, longitude) ; sst:scale_factor = 0.000552094668668839 ; sst:add_offset = 285.983000319853 ; sst:_FillValue = -32767s ; sst:missing_value = -32767s ; sst:units = "K" ; sst:long_name = "Sea surface temperature" ;

// global attributes: :Conventions = "CF-1.6" ; :history = "2017-08-04 06:17:58 GMT by grib_to_netcdf-2.4.0: grib_to_netcdf /data/data05/scratch/_mars-atls09-95e2cf679cd58ee9b4db4dd119a05a8d-gF5gxN.grib -o /data/data04/scratch/_grib2netcdf-atls01-a562cefde8a29a7288fa0b8b7f9413f7-VvH7PP.nc -utime" ; :_Format = "64-bit offset" ; } ``` Unfortunately removing the chunks didn't seem to help. I'm running with the pre-process workaround this morning to see if that completes. Sorry for the late response on this - been pretty busy.

I'll give both of those a shot.

For hosting, the files are currently on a local drive and they sum to about 1Tb. I can probably host a couple examples though.

Thanks again for the support.

That run was killed with the output

``bash ~/.pyenv/versions/3.4.6/lib/python3.4/site-packages/xarray/core/dtypes.py:23: FutureWarning: Conversion of the second argument of issubdtype fromfloattonp.floatingis deprecated. In future, it will be treated asnp.float64 == np.dtype(float).type`. if np.issubdtype(dtype, float):

Process finished with exit code 137 (interrupted by signal 9: SIGKILL) ```

I wasn't watching the machine at the time but I assume that's it falling over to memory pressure.

Hi @jhamman, I'm using 0.10.0 of xarray with dask 0.16.1 and distrobuted 1.18.0. I realise that last one is out of date, I will update and retry.

I'm just using whatever the default scheduler is as that's pretty much all the code I've got written above.

I'm unsure how to do a performance check as the dataset can't even be fully loaded currently. I've tried different chuck sizes in the past hoping to stumble on a magic size, but have been unsuccessful with that.

Sure, I'm running that now. I'll reply once/if it finished. Though watching my system monitor memory usage, it does not appear to be growing. I seem to remember the open function continually allocating itself more ram until it was killed.

I'll take a read through that issue while I wait.

That's true, maybe I misread last time or it's month dependant.

Hopefully this is what you're after - let me know if not. I used 3 *.nc files to make this, with the snippet you posted above.

bash <xarray.Dataset> Dimensions: (time: 728) Coordinates: longitude float32 10.0 latitude float32 10.0 * time (time) datetime64[ns] 1992-01-01 1992-01-01T03:00:00 ... Data variables: mwp (time) float64 dask.array<shape=(728,), chunksize=(127,)> Attributes: Conventions: CF-1.6 history: 2017-08-10 04:58:48 GMT by grib_to_netcdf-2.4.0: grib_to_ne... If you're after the entire dataset, I should be able to get that but may take some time.

Sure, this is the repr of a single file:

bash <xarray.Dataset> Dimensions: (time: 248) Coordinates: longitude float32 10.0 latitude float32 10.0 * time (time) datetime64[ns] 2004-12-01 2004-12-01T03:00:00 ... Data variables: mwd (time) float64 dask.array<shape=(248,), chunksize=(248,)> Attributes: Conventions: CF-1.6 history: 2017-08-09 16:22:56 GMT by grib_to_netcdf-2.4.0: grib_to_ne...

Thanks

Sorry to bump this. Still looking to a solution to this problem if anyone has had a similar experience. Thanks.

Hi @rabernat, thanks for the response. Sorry it's taken me a few days to get back to you.

Here's the info dump of one of the files:

``` xarray.Dataset { dimensions: latitude = 361 ; longitude = 720 ; time = 248 ;

variables: float32 longitude(longitude) ; longitude:units = degrees_east ; longitude:long_name = longitude ; float32 latitude(latitude) ; latitude:units = degrees_north ; latitude:long_name = latitude ; datetime64[ns] time(time) ; time:long_name = time ; float64 mwd(time, latitude, longitude) ; mwd:units = Degree true ; mwd:long_name = Mean wave direction ;

// global attributes: :Conventions = CF-1.6 ; :history = 2017-08-09 18:15:34 GMT by grib_to_netcdf-2.4.0: grib_to_netcdf /data/data05/scratch/_mars-atls02-70e05f9f8ba4e9d19932f1c45a7be8d8-Pwy6jZ.grib -o /data/data01/scratch/_grib2netcdf-atls02-95e2cf679cd58ee9b4db4dd119a05a8d-v4TKah.nc -utime ; ```

@shoyer great, thanks. I added the line below and it has reduced the size of the file down to that of the duplicate. Thanks pointing me the in the right direction. I'm assuming I do not need to fillnans with _FillValue after (though maybe I might).

python masked_ds.swh.encoding = {k: v for k, v in ds.swh.encoding.items() if k in {'_FillValue', 'add_offset', 'dtype', 'scale_factor'}}

@fmaussion @jhamman Ah great - that makes sense. I'll see if I can set them to the original file's short fill representation instead of nan.

Sure, here you go:

Original (128.9MB): ```bash $ ncdump -h -s swh_2010_01_05_05.nc netcdf swh_2010_01_05_05 { dimensions: longitude = 720 ; latitude = 361 ; time = UNLIMITED ; // (248 currently) variables: float longitude(longitude) ; longitude:units = "degrees_east" ; longitude:long_name = "longitude" ; float latitude(latitude) ; latitude:units = "degrees_north" ; latitude:long_name = "latitude" ; int time(time) ; time:units = "hours since 1900-01-01 00:00:0.0" ; time:long_name = "time" ; time:calendar = "gregorian" ; short swh(time, latitude, longitude) ; swh:scale_factor = 0.000203558072860934 ; swh:add_offset = 6.70098898894319 ; swh:_FillValue = -32767s ; swh:missing_value = -32767s ; swh:units = "m" ; swh:long_name = "Significant height of combined wind waves and swell" ;

// global attributes: :Conventions = "CF-1.6" ; :history = "2017-08-09 16:41:57 GMT by grib_to_netcdf-2.4.0: grib_to_netcdf /data/data04/scratch/_mars-atls01-a562cefde8a29a7288fa0b8b7f9413f7-5gV0xP.grib -o /data/data05/scratch/_grib2netcdf-atls09-70e05f9f8ba4e9d19932f1c45a7be8d8-jU8lEi.nc -utime" ; :_Format = "64-bit offset" ; } Duplicate (129.0MB):bash $ ncdump -h -s swh_2010_01_05_05-duplicate.nc netcdf swh_2010_01_05_05-duplicate { dimensions: longitude = 720 ; latitude = 361 ; time = UNLIMITED ; // (248 currently) variables: float longitude(longitude) ; longitude:_FillValue = NaNf ; longitude:units = "degrees_east" ; longitude:long_name = "longitude" ; longitude:_Storage = "contiguous" ; float latitude(latitude) ; latitude:_FillValue = NaNf ; latitude:units = "degrees_north" ; latitude:long_name = "latitude" ; latitude:_Storage = "contiguous" ; int time(time) ; time:long_name = "time" ; time:units = "hours since 1900-01-01" ; time:calendar = "gregorian" ; time:_Storage = "chunked" ; time:_ChunkSizes = 1024 ; time:_Endianness = "little" ; short swh(time, latitude, longitude) ; swh:_FillValue = -32767s ; swh:units = "m" ; swh:long_name = "Significant height of combined wind waves and swell" ; swh:add_offset = 6.70098898894319 ; swh:scale_factor = 0.000203558072860934 ; swh:_Storage = "chunked" ; swh:_ChunkSizes = 1, 361, 720 ; swh:_Endianness = "little" ;

// global attributes: :_NCProperties = "version=1|netcdflibversion=4.4.1.1|hdf5libversion=1.8.18" ; :Conventions = "CF-1.6" ; :history = "2017-08-09 16:41:57 GMT by grib_to_netcdf-2.4.0: grib_to_netcdf /data/data04/scratch/_mars-atls01-a562cefde8a29a7288fa0b8b7f9413f7-5gV0xP.grib -o /data/data05/scratch/_grib2netcdf-atls09-70e05f9f8ba4e9d19932f1c45a7be8d8-jU8lEi.nc -utime" ; :_Format = "netCDF-4" ; } Masked (515.7MB):bash $ ncdump -h -s swh_2010_01_05_05-masked.nc netcdf swh_2010_01_05_05-masked { dimensions: longitude = 720 ; latitude = 361 ; time = 248 ; variables: float longitude(longitude) ; longitude:_FillValue = NaNf ; longitude:units = "degrees_east" ; longitude:long_name = "longitude" ; longitude:_Storage = "contiguous" ; float latitude(latitude) ; latitude:_FillValue = NaNf ; latitude:units = "degrees_north" ; latitude:long_name = "latitude" ; latitude:_Storage = "contiguous" ; int time(time) ; time:long_name = "time" ; time:units = "hours since 1900-01-01" ; time:calendar = "gregorian" ; time:_Storage = "contiguous" ; time:_Endianness = "little" ; double swh(time, latitude, longitude) ; swh:_FillValue = NaN ; swh:units = "m" ; swh:long_name = "Significant height of combined wind waves and swell" ; swh:_Storage = "contiguous" ;

// global attributes: :_NCProperties = "version=1|netcdflibversion=4.4.1.1|hdf5libversion=1.8.18" ; :Conventions = "CF-1.6" ; :history = "2017-08-09 16:41:57 GMT by grib_to_netcdf-2.4.0: grib_to_netcdf /data/data04/scratch/_mars-atls01-a562cefde8a29a7288fa0b8b7f9413f7-5gV0xP.grib -o /data/data05/scratch/_grib2netcdf-atls09-70e05f9f8ba4e9d19932f1c45a7be8d8-jU8lEi.nc -utime" ; :_Format = "netCDF-4" ; } ``` I assume it's about that fill/missing value changing? Thanks for the help.

Increasing the chunk sizes seemed to resolve this issue. I was loading readings over time on the world as 360x180xT was tryin to rechunk them to 1x1xT.