Open jsadler2 opened 4 years ago
Hi @jsadler2 - Can you show us what sm_from_zarr
looks like (print(sm_from_zarr)
) will do.
Sure
In [5]: print(ds)
<xarray.Dataset>
Dimensions: (datetime: 20, site_code: 20)
Coordinates:
* datetime (datetime) datetime64[ns] 1970-01-01 ... 1970-01-01T19:00:00
* site_code (site_code) object '01302020' '01303000' ... '01315000'
Data variables:
streamflow (datetime, site_code) float64 dask.array<shape=(20, 20), chunksize=(20, 20)>
Thanks @jsadler2 - I think this is a bug in xarray. We should be able to round trip the site_coordinate
variable.
I'm experiencing the same issue, which seems to be also related to one of my coordinates having object as datatype. Luckily, the workaround proposed by @jsadler2 works in my case, too.
I ran into the same issue. It seems like zarr is inserting VLenUTF8
as a filter, but the loaded data array already has that as a filter so it's trying to double encode . So another workaround is to delete the filters on site_code
:
import xarray as xr
sm_from_zarr = xr.open_zarr('tmp/test_sm_zarr')
del sm_from_zarr.site_code.encoding["filters"]
sm_from_zarr.to_zarr('tmp/test_sm_zarr_from', mode='w')
I ran into the same issue. It seems like zarr is inserting VLenUTF8 as a filter, but the loaded data array already has that as a filter so it's trying to double encode
Indeed. And it appears that these lines are the culprits:
ipdb> v
<xarray.Variable (x: 3)>
array(['a', 'b', 'c'], dtype=object)
ipdb> check
False
ipdb> vn
'x'
ipdb> encoding = extract_zarr_variable_encoding(v, raise_on_invalid=check, name=vn)
ipdb> encoding
{'chunks': (3,), 'compressor': Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0), 'filters': [VLenUTF8()]}
Zarr appears to be ignoring the filter information from xarray. Zarr proceeds to extracting its own filter. As a result, we end up with two filters:
ipdb> zarr_array = self.ds.create(name, shape=shape, dtype=dtype, fill_value=fill_value, **encoding)
ipdb> self.ds['x']._meta
{'zarr_format': 2, 'shape': (3,), 'chunks': (3,), 'dtype': dtype('O'), 'compressor': {'blocksize': 0, 'clevel': 5, 'cname': 'lz4', 'id': 'blosc', 'shuffle': 1}, 'fill_value': None, 'order': 'C', 'filters': [{'id': 'vlen-utf8'}, {'id': 'vlen-utf8'}]}
ipdb> self.ds['x']._meta['filters']
[{'id': 'vlen-utf8'}, {'id': 'vlen-utf8'}]
As @borispf and @jsadler2 suggested, clearing the filters from encoding before initiating the zarr store creation works:
ipdb> enc # without filters
{'chunks': (3,), 'compressor': Blosc(cname='lz4', clevel=5, shuffle=SHUFFLE, blocksize=0), 'filters': []}
ipdb> zarr_array = self.ds.create(name, shape=shape, dtype=dtype, fill_value=fill_value, **enc)
ipdb> self.ds['x']._meta['filters']
[{'id': 'vlen-utf8'}]
@jhamman since you are more familiar with the internals of zarr + xarray, should we default to ignoring filter information from xarray and let zarr take care of the extraction of filter information?
Thanks @andersy005 for the write up and for digging into this. Doesn't this seem like it could be a bug in Zarr's create
method?
Hi, I also keep running into this issue all the time.
Right now, there is no way of round-tripping xr.open_zarr().to_zarr()
, also because of https://github.com/pydata/xarray/issues/5219.
The only workaround that seems to help is the following:
to_store = xrds.copy()
for var in to_store.variables:
to_store[var].encoding.clear()
Thanks for that, looks like I just came across this.
Think it worked with one variable - @Hoeze workaround was necessary for more than one.
This has been happening a lot to me lately when writing to zarr. Thanks to @bolliger32 for the tip - this usually works like a charm:
for v in list(ds.coords.keys()):
if ds.coords[v].dtype == object:
ds.coords[v] = ds.coords[v].astype("unicode")
for v in list(ds.variables.keys()):
if ds[v].dtype == object:
ds[v] = ds[v].astype("unicode")
For whatever reason, clearing the encoding and/or using .astype(str)
doesn't seem to work as reliably. I don't have a good MRE for this but hope it helps others with the same issue.
note the flag raised by @FlorisCalkoen below - don't just throw this at all your writes! there are other object types (e.g. CFTime) which you probably don't want to convert to string. This is just a patch to get around this issue for dataarrays with string coords/variables.
@delgadom I just noticed that your proposed solution has the side effect of also converting cftime
objects (e.g., below) to unicode strings.
xarray.DataArray 'time' (time: 1) array([cftime.DatetimeNoLeap(2007, 7, 2, 12, 0, 0, 0, has_year_zero=True)], dtype=object)
I updated your lines using @Hoeze' clear function and that seems to work for now.
for v in list(ds.coords.keys()):
if ds.coords[v].dtype == object:
ds[v].encoding.clear()
for v in list(ds.variables.keys()):
if ds[v].dtype == object:
ds[v].encoding.clear()
ha - yeah that's a good flag. I definitely didn't intend for that to be a universally applied patch! so probably should have included a buyer beware. but we did find that clearing the encoding doesn't always do the trick for string arrays. So a comprehensive patch will probably need to be more nuanced.
My impression is that keeping the zarr encoding leads to a bunch of issues (see my issue above) or the current one. There also seems to be an issue with chunking being preserved because the array encodings arent overwritten, but I cant find that issue right now. Since all of these issues are resolved by popping the zarr encoding I am wondering what are the downsides of this and whether it'd be easier to not keep that encoding at all?
I find this annoying too!
I think we can fold it into https://github.com/pydata/xarray/issues/6323, though it's also common enough that maybe we leave it open for future travellers
I hit this issue this week, and the workaround in https://github.com/pydata/xarray/issues/3476#issuecomment-1115045538 worked for me.
I run into this issue fairly commonly (usually during playing copying of a dataset xr.open_zarr(..).to_zarr(..)
) and the fix in https://github.com/pydata/xarray/issues/3476#issuecomment-1115045538 works for me.
I saved an xarray dataset to zarr using
to_zarr
. I then later tried to read that dataset from the original zarr, re-chunk it, and then write to a new zarr. When I did that I get a strange error. I attached a zip of minimal version of the zarr dataset that I am using. test_sm_zarr.zipMCVE Code Sample
Expected Output
No error
Problem Description
I get this error:
BUT I think it has something to do with the datatype of one of my coordinates,
site_code
. Because, if it do this I get no error:Before converting the datatype of the
site_code
coordinate isobject
Output of
xr.show_versions()