Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Problems faced while storing onto Zarr store using ABSStore #528

Open
dokooh opened this issue Dec 12, 2019 · 25 comments
Open

Problems faced while storing onto Zarr store using ABSStore #528

dokooh opened this issue Dec 12, 2019 · 25 comments

Comments

@dokooh
Copy link

dokooh commented Dec 12, 2019 •
edited
Loading 

# Your code here

import zarr
from azure.storage.blob import BlockBlobService

store = zarr.ABSStore(container='zarrstoreall', prefix='zarrstoreall',account_name='xxxx',account_key='xxxx', blob_service_kwargs={'is_emulated': False})

compressor = zarr.Blosc(cname='zstd', clevel=3)
encoding = {vname: {'compressor': compressor} for vname in ds.data_vars}
ds.to_zarr(store=store, encoding=encoding, consolidated=True)

Problem description

I'm trying to use ABSStore to store a large XArray dataset onto a zarr store using blob store. (see the code in previous section). I am facing two issues currently:

  1. I am getting first some sort of network error when loading "certain" variables into the store:
    image

After some time passing I get this error:
image

Needless to say with relatively smaller sizes of XArray datasets I did not face these issues.

I appreciate your kind attention.

Version and installation information

Please provide the following:

  • Value of zarr.__version__ = '2.3.2'
  • Value of numcodecs.__version__ = '0.6.4'
  • Version of Python interpreter = Python 3.7.3
  • Operating system (Linux/Windows/Mac) = Databricks Runtime Version
    6.1 (includes Apache Spark 2.4.4, Scala 2.11)
  • How Zarr was installed (e.g., "using pip into virtual environment", or "using conda")
    !pip install zarr

Also, if you think it might be relevant, please provide the output from pip freeze or
conda env export depending on which was used to install Zarr.
pip freeze output:
adal==1.2.2
asciitree==0.3.3
asn1crypto==0.24.0
azure==4.0.0
azure-applicationinsights==0.1.0
azure-batch==4.1.3
azure-common==1.1.23
azure-cosmosdb-nspkg==2.0.2
azure-cosmosdb-table==1.0.6
azure-datalake-store==0.0.48
azure-eventgrid==1.3.0
azure-graphrbac==0.40.0
azure-keyvault==1.1.0
azure-loganalytics==0.1.0
azure-mgmt==4.0.0
azure-mgmt-advisor==1.0.1
azure-mgmt-applicationinsights==0.1.1
azure-mgmt-authorization==0.50.0
azure-mgmt-batch==5.0.1
azure-mgmt-batchai==2.0.0
azure-mgmt-billing==0.2.0
azure-mgmt-cdn==3.1.0
azure-mgmt-cognitiveservices==3.0.0
azure-mgmt-commerce==1.0.1
azure-mgmt-compute==4.6.2
azure-mgmt-consumption==2.0.0
azure-mgmt-containerinstance==1.5.0
azure-mgmt-containerregistry==2.8.0
azure-mgmt-containerservice==4.4.0
azure-mgmt-cosmosdb==0.4.1
azure-mgmt-datafactory==0.6.0
azure-mgmt-datalake-analytics==0.6.0
azure-mgmt-datalake-nspkg==3.0.1
azure-mgmt-datalake-store==0.5.0
azure-mgmt-datamigration==1.0.0
azure-mgmt-devspaces==0.1.0
azure-mgmt-devtestlabs==2.2.0
azure-mgmt-dns==2.1.0
azure-mgmt-eventgrid==1.0.0
azure-mgmt-eventhub==2.6.0
azure-mgmt-hanaonazure==0.1.1
azure-mgmt-iotcentral==0.1.0
azure-mgmt-iothub==0.5.0
azure-mgmt-iothubprovisioningservices==0.2.0
azure-mgmt-keyvault==1.1.0
azure-mgmt-loganalytics==0.2.0
azure-mgmt-logic==3.0.0
azure-mgmt-machinelearningcompute==0.4.1
azure-mgmt-managementgroups==0.1.0
azure-mgmt-managementpartner==0.1.1
azure-mgmt-maps==0.1.0
azure-mgmt-marketplaceordering==0.1.0
azure-mgmt-media==1.0.0
azure-mgmt-monitor==0.5.2
azure-mgmt-msi==0.2.0
azure-mgmt-network==2.7.0
azure-mgmt-notificationhubs==2.1.0
azure-mgmt-nspkg==3.0.2
azure-mgmt-policyinsights==0.1.0
azure-mgmt-powerbiembedded==2.0.0
azure-mgmt-rdbms==1.9.0
azure-mgmt-recoveryservices==0.3.0
azure-mgmt-recoveryservicesbackup==0.3.0
azure-mgmt-redis==5.0.0
azure-mgmt-relay==0.1.0
azure-mgmt-reservations==0.2.1
azure-mgmt-resource==2.2.0
azure-mgmt-scheduler==2.0.0
azure-mgmt-search==2.1.0
azure-mgmt-servicebus==0.5.3
azure-mgmt-servicefabric==0.2.0
azure-mgmt-signalr==0.1.1
azure-mgmt-sql==0.9.1
azure-mgmt-storage==2.0.0
azure-mgmt-subscription==0.2.0
azure-mgmt-trafficmanager==0.50.0
azure-mgmt-web==0.35.0
azure-nspkg==3.0.2
azure-servicebus==0.21.1
azure-servicefabric==6.3.0.0
azure-servicemanagement-legacy==0.20.6
azure-storage-blob==1.5.0
azure-storage-common==1.4.2
azure-storage-file==1.4.0
azure-storage-queue==1.4.0
backcall==0.1.0
boto==2.49.0
boto3==1.9.162
botocore==1.12.163
certifi==2019.3.9
cffi==1.12.2
cftime==1.0.4.2
chardet==3.0.4
cryptography==2.6.1
cycler==0.10.0
Cython==0.29.6
dask==2.9.0
decorator==4.4.0
docutils==0.14
fasteners==0.15
fsspec==0.6.1
idna==2.8
ipython==7.4.0
ipython-genutils==0.2.0
isodate==0.6.0
jedi==0.13.3
jmespath==0.9.4
kiwisolver==1.1.0
koalas==0.23.0
locket==0.2.0
matplotlib==3.0.3
monotonic==1.5
msrest==0.6.10
msrestazure==0.6.2
netCDF4==1.5.3
numcodecs==0.6.4
numpy==1.16.2
oauthlib==3.1.0
pandas==0.24.2
parso==0.3.4
partd==1.1.0
patsy==0.5.1
pexpect==4.6.0
pickleshare==0.7.5
prompt-toolkit==2.0.9
psycopg2==2.7.6.1
ptyprocess==0.6.0
pyarrow==0.13.0
pycparser==2.19
pycurl==7.43.0
Pygments==2.3.1
pygobject==3.20.0
PyJWT==1.7.1
pyOpenSSL==19.0.0
pyparsing==2.4.2
PySocks==1.6.8
python-apt==1.1.0b1+ubuntu0.16.4.5
python-dateutil==2.8.0
pytz==2018.9
requests==2.21.0
requests-oauthlib==1.3.0
s3transfer==0.2.1
scikit-learn==0.20.3
scipy==1.2.1
seaborn==0.9.0
six==1.12.0
ssh-import-id==5.5
statsmodels==0.9.0
toolz==0.10.0
traitlets==4.3.2
unattended-upgrades==0.1
urllib3==1.24.1
virtualenv==16.4.1
wcwidth==0.1.7
xarray==0.14.1
zarr==2.3.2
@alimanfoo
Copy link
Member

alimanfoo commented Dec 12, 2019 via email

@tjcrone
Copy link
Member

tjcrone commented Dec 12, 2019

I believe the first error is actually a warning, and occurs when Zarr looks for metadata files that do not exist. This has been solved in newer versions of the Azure SDK. I would try upgrading azure-storage-blob to v2.1.

It's worth noting that while investigating this I learned that there is a major new release of the Azure SDK that looks like it will break ABSStore entirely. We are going to need to figure out how to deal with this probably soon. It's not obvious how we are going to deal with two versions of the SDK that are essentially incompatible. I will probably start a new issue to work on this eventually.

@dokooh
Copy link
Author

dokooh commented Dec 12, 2019

Thanks @alimanfoo I use mfdataset method and I get some warnings during the import which should suggest that:

/local_disk0/tmp/1576146109393-0/PythonShell.py:4: FutureWarning: In xarray version 0.15 the default behaviour of open_mfdataset
will change. To retain the existing behavior, pass
combine='nested'. To use future default behavior, pass
combine='by_coords'. See
http://xarray.pydata.org/en/stable/combining.html#combining-multi

import errno
/databricks/python/lib/python3.7/site-packages/xarray/backends/api.py:933: FutureWarning: The datasets supplied have global dimension coordinates. You may want
to use the new combine_by_coords function (or the
combine='by_coords' option to open_mfdataset) to order the datasets
before concatenation. Alternatively, to continue concatenating based
on the order the datasets are supplied in future, please use the new
combine_nested function (or the combine='nested' option to
open_mfdataset). from_openmfds=True,

Would this suggest that some of the files were not loaded I guess into Xarray. I will try experimenting with the combine options to check this.

@jakirkham
Copy link
Member

Maybe not related, but did you see PR ( #526 )?

@shikharsg
Copy link
Member

shikharsg commented Dec 12, 2019 •
edited
Loading

I believe the first error is actually a warning, and occurs when Zarr looks for metadata files that do not exist. This has been solved in newer versions of the Azure SDK. I would try upgrading azure-storage-blob to v2.1.

It's worth noting that while investigating this I learned that there is a major new release of the Azure SDK that looks like it will break ABSStore entirely. We are going to need to figure out how to deal with this probably soon. It's not obvious how we are going to deal with two versions of the SDK that are essentially incompatible. I will probably start a new issue to work on this eventually.

Agree with @tjcrone here, the first warning goes away after updating to the newer version.

As for the above error, I have faced various errors, mostly out of memory error(so it's worth monitoring the memory of your device/vm while doing the above) but also the one above while transferring large amounts of netCDF data to zarr. My solution was to transfer the data to zarr "in parts". It is easily possible now with xarray's new "append" feature for zarr. You can use ds.to_zarr with mode='a' and also provide the dimension along which the data will be appended. See here: http://xarray.pydata.org/en/stable/generated/xarray.Dataset.to_zarr.html or here: pydata/xarray#2706

@dokooh
Copy link
Author

dokooh commented Jan 9, 2020

Hi,
we did some investigation and realized that we have some NaN values in the dataset which we were not aware of. 300 rows in total in a full year chunk. Could all this be caused by NaN/Inf values?

@alimanfoo
Copy link
Member

Hi,
we did some investigation and realized that we have some NaN values in the dataset which we were not aware of. 300 rows in total in a full year chunk. Could all this be caused by NaN/Inf values?

I would not have thought so, at least on the side of writing the zarr data, zarr should be ignorant to what the actual data values are, it will just write them.

But it's still unclear to me at least whether the errors are being generated during the read from netcdf or the write to zarr. The error messages suggest it's the read from netcdf that's triggering the error, but I may have misunderstood. Are you reading the netcdf data from ABS, or is that being read from a local file system? Apologies if I'm barking up the wrong tree.

@dokooh
Copy link
Author

dokooh commented Jan 10, 2020 •
edited
Loading

Hi,
we did some investigation and realized that we have some NaN values in the dataset which we were not aware of. 300 rows in total in a full year chunk. Could all this be caused by NaN/Inf values?

I would not have thought so, at least on the side of writing the zarr data, zarr should be ignorant to what the actual data values are, it will just write them.

But it's still unclear to me at least whether the errors are being generated during the read from netcdf or the write to zarr. The error messages suggest it's the read from netcdf that's triggering the error, but I may have misunderstood. Are you reading the netcdf data from ABS, or is that being read from a local file system? Apologies if I'm barking up the wrong tree.

Thanks for your kind follow up. We are reading NetCDF from local file system through Xray and then writing it onto Zarr.

@shikharsg
Copy link
Member

Re: getting NaN values

I think I might have found out why this happens, as I ran into this myself.

There is a fill_value attribute in zarr, which zarr uses to fill out missing chunks (see here).

Xarray uses this same attribute as the _FillValue attribute(see here) for decoding using the CF conventions, which is something quite different from filling out missing chunks.

@zarr-developers/core-devs Is this a correct interpretation? If so where should this be fixed? In xarray or in zarr?

@dokooh I fixed this temporarily by giving mask_and_scale=False to xr.open_zarr

@alimanfoo
Copy link
Member

Hi @shikharsg, thanks a lot for following up.

Yes zarr has a fill_value attribute in the array metadata, and this is used to fill out missing chunks.

I don't know the details of how xarray zarr backend uses the _FillValue attribute, defer to @rabernat and @jhamman.

I'm still not sure what the underlying problem is here. @shikharsg do you have a handle on where the problem is? Could you elaborate?

@martindurant
Copy link
Member

If you are curious, would you be willing to trial abfs via fsspec, as allowed by #546 (implements only for nested store for now, so may not work for you) ? You would need https://github.com/dask/adlfs

@shikharsg
Copy link
Member

shikharsg commented Mar 26, 2020 •
edited
Loading

So I had a large number of netCDF files which i transferred to zarr, back in October 2018. This was when xr.to_zarr with the append feature did not exist. So when it was finally released sometime mid last year, I had to manually build up the _ARRAY_DIMENSIONS attribute, as I have done below in the small reproducible example.

Python 3.7.6 (default, Jan  8 2020, 19:59:22) 
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import zarr
>>> import xarray as xr
>>> import numpy as np
>>> zarr.__version__, xr.__version__, np.__version__
('2.4.0', '0.15.0', '1.18.1')
>>>
>>> # in memory zarr array
>>> store = zarr.MemoryStore()
>>> grp = zarr.open_group(store)
>>> arr = zarr.open_array(store, path='foo', shape=(2, 10), fill_value=0.0, chunks=(1, 10))
>>> arr[0] = np.zeros((10,))
>>> arr[0] = np.ones((10,))
>>> 
>>> arr[:]
array([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])
>>>
>>> # manually build up dimensions
>>> dim1 = zarr.open_array(store, path='dim1', shape=(2,))
>>> dim1[:] = np.array(list(range(1, 3)))
>>> dim1.attrs['_ARRAY_DIMENSIONS'] = ['dim1']
>>> dim2 = zarr.open_array(store, path='dim2', shape=(10,))
>>> dim2[:] = np.array(list(range(1, 11)))
>>> dim2.attrs['_ARRAY_DIMENSIONS'] = ['dim2']
>>> arr.attrs['_ARRAY_DIMENSIONS'] = ['dim1', 'dim2']
>>> 
>>> xr.open_zarr(store)['foo'].values
array([[ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.],
       [nan, nan, nan, nan, nan, nan, nan, nan, nan, nan]])
>>> 
>>> arr[:]
array([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])

As you can see, zarr and xarray return different results. This is because xarray uses the fill_value attribute to replace all 0 values with nan

@shikharsg
Copy link
Member

Perhaps this is a more appropriate example, where you can see zarr and xarray are using fill_value in different ways:

>>> zarr.__version__, xr.__version__, np.__version__
('2.4.0', '0.15.0', '1.18.1')
>>> 
>>> # in memory zarr array
>>> store = zarr.MemoryStore()
>>> grp = zarr.open_group(store)
>>> arr = zarr.open_array(store, path='foo', shape=(2, 10), fill_value=0.0, chunks=(1, 10))
>>> arr[0] = np.ones((10,))
>>> 
>>> arr[:]
array([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])
>>> 
>>> # manually build up dimensions
>>> dim1 = zarr.open_array(store, path='dim1', shape=(2,))
>>> dim1[:] = np.array(list(range(1, 3)))
>>> dim1.attrs['_ARRAY_DIMENSIONS'] = ['dim1']
>>> dim2 = zarr.open_array(store, path='dim2', shape=(10,))
>>> dim2[:] = np.array(list(range(1, 11)))
>>> dim2.attrs['_ARRAY_DIMENSIONS'] = ['dim2']
>>> arr.attrs['_ARRAY_DIMENSIONS'] = ['dim1', 'dim2']
>>> 
>>> print(xr.open_zarr(store)['foo'].values)
[[ 1.  1.  1.  1.  1.  1.  1.  1.  1.  1.]
 [nan nan nan nan nan nan nan nan nan nan]]
>>>

@dazzag24
Copy link

If you are curious, would you be willing to trial abfs via fsspec, as allowed by #546 (implements only for nested store for now, so may not work for you) ? You would need https://github.com/dask/adlfs

@martindurant Is the adlfs work appropriate for files stored in standard Azure blob storage? From the description it looks like it targets the datalake storage?
Thanks

@martindurant
Copy link
Member

It implements both datalake and blob. The latter is more recent, but I believe it is complete.

@shikharsg
Copy link
Member

@martindurant is this in context to the current issue or just in general?

@martindurant
Copy link
Member

In hindsight, it probably makes no difference to how the nan-value is inferred by zarr versus xarray ; so in general.

@shikharsg
Copy link
Member

Would love to try it. Will try to check it out over the next couple of days.

@dazzag24
Copy link

@martindurant Does it support SAS tokens? I see the example mentions only

STORAGE_OPTIONS={'account_name': ACCOUNT_NAME, 'account_key': ACCOUNT_KEY}

@martindurant
Copy link
Member

I have no idea what SAS tokens are :|
You would have to ask in an issue at adlfs. Perhaps this was a red herring!
Alternatively, if zarr's ABS store supports this auth and adlfs does not, it probably would be trivial to port the code.

@alimanfoo alimanfoo mentioned this issue Mar 27, 2020
Closed
@shikharsg
Copy link
Member

So I had a large number of netCDF files which i transferred to zarr, back in October 2018. This was when xr.to_zarr with the append feature did not exist. So when it was finally released sometime mid last year, I had to manually build up the _ARRAY_DIMENSIONS attribute, as I have done below in the small reproducible example.

Python 3.7.6 (default, Jan  8 2020, 19:59:22) 
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import zarr
>>> import xarray as xr
>>> import numpy as np
>>> zarr.__version__, xr.__version__, np.__version__
('2.4.0', '0.15.0', '1.18.1')
>>>
>>> # in memory zarr array
>>> store = zarr.MemoryStore()
>>> grp = zarr.open_group(store)
>>> arr = zarr.open_array(store, path='foo', shape=(2, 10), fill_value=0.0, chunks=(1, 10))
>>> arr[0] = np.zeros((10,))
>>> arr[0] = np.ones((10,))
>>> 
>>> arr[:]
array([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])
>>>
>>> # manually build up dimensions
>>> dim1 = zarr.open_array(store, path='dim1', shape=(2,))
>>> dim1[:] = np.array(list(range(1, 3)))
>>> dim1.attrs['_ARRAY_DIMENSIONS'] = ['dim1']
>>> dim2 = zarr.open_array(store, path='dim2', shape=(10,))
>>> dim2[:] = np.array(list(range(1, 11)))
>>> dim2.attrs['_ARRAY_DIMENSIONS'] = ['dim2']
>>> arr.attrs['_ARRAY_DIMENSIONS'] = ['dim1', 'dim2']
>>> 
>>> xr.open_zarr(store)['foo'].values
array([[ 1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.,  1.],
       [nan, nan, nan, nan, nan, nan, nan, nan, nan, nan]])
>>> 
>>> arr[:]
array([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]])

As you can see, zarr and xarray return different results. This is because xarray uses the fill_value attribute to replace all 0 values with nan

@rabernat @jhamman would be great to have your comments on this

@dazzag24
Copy link

I have no idea what SAS tokens are :|

@martindurant FYI SAS tokens are a way of allowing access to your blob stores with more fine grained and potentially time limited access
https://docs.microsoft.com/en-us/azure/storage/common/storage-sas-overview

For example you could give someone read-only access for period of one month via a token.

@martindurant
Copy link
Member

OK, so some sort of delegation thing...
Still, the conversation of how to use these with the abfs fsspec backend, if it doesn't already should happen in the adlfs repo, pointing to the one in zarr, if that already works. I am just trying to rationalise the number of places that storage things are defined...

@Chroxvi
Copy link

Chroxvi commented Jul 13, 2020

It's worth noting that while investigating this I learned that there is a major new release of the Azure SDK that looks like it will break ABSStore entirely. We are going to need to figure out how to deal with this probably soon. It's not obvious how we are going to deal with two versions of the SDK that are essentially incompatible. I will probably start a new issue to work on this eventually

@tjcrone Did you ever create a new issue for this? I can't seem to find one. Unfortunately, the version 12 of the azure-storage-blob SDK does break ABSStore entirely.

@jhamman jhamman mentioned this issue Sep 24, 2020
Closed
@rabernat rabernat mentioned this issue May 3, 2021
Open
@jakirkham
Copy link
Member

cc @TomAugspurger (who may have thoughts here 🙂)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
8 participants
@dokooh @alimanfoo @dazzag24 @jakirkham @Chroxvi @tjcrone @martindurant @shikharsg