issue_comments

This might be worth testing with the changes from https://github.com/pydata/xarray/pull/2261, which refactors xarray's IO handling.

I updated my example above to show that the chunking over the last dimension is ridiculously slow.

Sorry for the confusion, I had an obvious mistake in my timing experiment above (forgot to do the actual computations...).

The dimension order does make a difference:

```python import dask as da import xarray as xr

d = xr.DataArray(da.array.zeros((1000, 721, 1440), chunks=(10, 721, 1440)), dims=('z', 'y', 'x')) d.to_netcdf('da.nc') # 8.3 Gb

with xr.open_dataarray('da.nc', chunks={'z':10}) as d: %timeit d.sum().load() 3.94 s ± 95.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

with xr.open_dataarray('da.nc', chunks={'y':10}) as d: %timeit d.sum().load() 4.15 s ± 316 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

with xr.open_dataarray('da.nc', chunks={'x':10}) as d: %timeit d.sum().load() 1min 54s ± 1.43 s per loop (mean ± std. dev. of 7 runs, 1 loop each)

with xr.open_dataarray('da.nc', chunks={'y':10, 'x':10}) as d: %timeit d.sum().load() 2min 23s ± 215 ms per loop (mean ± std. dev. of 7 runs, 1 loop each) ```

Those chunksizes are the opposite of what I was expecting...

chunksizes in encoding are ignored in your case, dask still uses your user provided encoding.

Can you still try to chunk along one dimension only? i.e. chunks={'time':200}

[EDIT]: forgot the load ...

<s> forget my comment about chunks - I thought this would make a difference but it's actually the opposite (to my surprise): </s>

Out of curiosity: - why do you chunk over lats and lons rather than time? The order of dimensions in your dataarray suggest that chunking over time could be more efficient - can you show the output of ds.mtpr and ds.mtpr.encoding ?

Can you forget about zarr for a moment and just do a reduction on your dataset? For example: python ds.sum().load()

Keep the same chunk arguments you are currently using. This will help us understand if the problem is with reading the files.

Is it your intention to chunk the files contiguously in time? Depending on the underlying structure of the data within the netCDF file, this could amount to a complete transposition of the data, which could be very slow / expensive. This could have some parallels with #2004.

I was somehow expecting that each worker will read a chunk and then write it to zarr, streamlined.

Yes, this is what we want!

@lrntct - this sounds like a reasonable way to use zarr. We routinely do this sort of transcoding and it works reasonable well. Unfortunately something clearly isn't working right in your case. These things can be hard to debug, but we will try to help you.

You might want to start by reviewing the guide I wrote for Pangeo on preparing zarr datasets.

It would also be good to see a bit more detail. You posted a function netcdf2zarr that converts a single netcdf file to a single zarr file. How are you invoking that function? Are you trying to create one zarr store for each netCDF file? How many netCDF files are there? If there are many (e.g. one per timmestep), my recommendation is to create only one zarr store for the whole dataset. Open the netcdf files using open_mfdataset.

If instead you have just one big netCDF file as in the example you posted above, I think I see you problem: you are calling .chunk() after calling open_dataset(), rather calling open_dataset(nc_path, chunks=chunks). This probably means that you are loading the whole dataset in a single task and then re-chunking. That could be the source of the inefficiency.

More ideas: - explicitly specify the chunks (rather than using 'auto') - eliminate the negative number in your chunk sizes - make sure you really need clevel=9

Another useful piece of advice would be to use the dask distributed dashboard to monitor what is happening under the hood. You can do this by running python from dask.distributed import Client client = Client() client In a notebook, this should provide you a link to the scheduler dashboard. Once you call ds.to_zarr(), watch the task stream in the dashboard to see what is happening.

Hopefully these ideas can help you move forward.