dask.async.RuntimeError: NetCDF: HDF error on xarray to_netcdf

[pwolfram]

Dask appears to be failing on serialization following a ds.to_netcdef() via a NETCDF: HDF error.
Excerpted error below:

Traceback (most recent call last):
  File "reduce_dispersion_file.py", line 40, in <module>
    if __name__ == "__main__":
  File "reduce_dispersion_file.py", line 36, in reduce_dispersion_file
    with timeit_context('output to disk'):
  File "/users/pwolfram/envs/LIGHT_analysis/lib/python2.7/site-packages/xarray/core/dataset.py", line 791, in to_netcdf
    engine=engine, encoding=encoding)
  File "/users/pwolfram/envs/LIGHT_analysis/lib/python2.7/site-packages/xarray/backends/api.py", line 356, in to_netcdf
    dataset.dump_to_store(store, sync=sync, encoding=encoding)
  File "/users/pwolfram/envs/LIGHT_analysis/lib/python2.7/site-packages/xarray/core/dataset.py", line 739, in dump_to_store
    store.sync()
  File "/users/pwolfram/envs/LIGHT_analysis/lib/python2.7/site-packages/xarray/backends/netCDF4_.py", line 283, in sync
    super(NetCDF4DataStore, self).sync()
  File "/users/pwolfram/envs/LIGHT_analysis/lib/python2.7/site-packages/xarray/backends/common.py", line 186, in sync
    self.writer.sync()
  File "/users/pwolfram/envs/LIGHT_analysis/lib/python2.7/site-packages/xarray/backends/common.py", line 165, in sync
    da.store(self.sources, self.targets)
  File "/users/pwolfram/lib/python2.7/site-packages/dask/array/core.py", line 712, in store
    Array._get(dsk, keys, **kwargs)
  File "/users/pwolfram/lib/python2.7/site-packages/dask/base.py", line 43, in _get
    return get(dsk2, keys, **kwargs)
  File "/users/pwolfram/lib/python2.7/site-packages/dask/threaded.py", line 57, in get 
    **kwargs)
  File "/users/pwolfram/lib/python2.7/site-packages/dask/async.py", line 481, in get_async
    raise(remote_exception(res, tb))
dask.async.RuntimeError: NetCDF: HDF error

Traceback
---------
  File "/users/pwolfram/lib/python2.7/site-packages/dask/async.py", line 264, in execute_task
    result = _execute_task(task, data)
  File "/users/pwolfram/lib/python2.7/site-packages/dask/async.py", line 246, in _execute_task
    return func(*args2)
  File "/users/pwolfram/lib/python2.7/site-packages/dask/array/core.py", line 1954, in store
    out[index] = np.asanyarray(x)
  File "netCDF4/_netCDF4.pyx", line 3678, in netCDF4._netCDF4.Variable.__setitem__ (netCDF4/_netCDF4.c:37215)
  File "netCDF4/_netCDF4.pyx", line 3887, in netCDF4._netCDF4.Variable._put (netCDF4/_netCDF4.c:38907)

Script used: https://gist.github.com/98acaa31a4533b490f78
Full output: https://gist.github.com/248efce774ad08cb1dd6

[pwolfram]

cc @mrocklin

[pwolfram]

I should note that serialization also does not appear to be robust under reshaping the data via ds = ds.transpose('Nt','Nt-1','Nr','Nb','Nc') as well as rechunking. The input data stream was previously generated via a call to ds.to_netcdf in another script using xarray.

[pwolfram]

There are a large number of files (1320) where nfiles = 120 and len(dslist)=11, so perhaps this is an issue with opening a large number of files as noted by @rabernat.

[mrocklin]

1024 might be a common open file handle limit. Some things to try to isolate the issue:

1. Try this with dask.set_globals(get=dask.async.get_sync) to turn off threading
2. Try just opening all of the files and see if the NetCDF error presents itself under normal operation

[pwolfram]

Quick question @mrocklin, for 2, are you proposing a script that just opens all the files, e.g., something like this

# get full xr dataset
dslist = []
nfiles = len(glob.glob('dispersion_calcs_rlzn0*layerrange_0000-0000.nc'))
for i in np.arange(nfiles):                                                                                                                        
  ds = xr.open_mfdataset('dispersion_calcs_rlzn%04d_*nc'%(i))
  dslist.append(ds)
dstotal = xr.concat(dslist,'Nr')
# do an operation spanning Nr space and Nb space
print dstotal.dtdays.values

where dtdays spans the all the files? I'm running it now.

[mrocklin]

Sure. I'm not proposing any particular approach. I'm just supporting your previous idea that maybe the problem is having too many open file handles. It would be good to check this before diving into threading or concurrency issues.

[pwolfram]

Agreed. I'll let you know what I find out. Thanks @mrocklin.

[pwolfram]

Test 2 passed, so it doesn't appear to be due to too many open file handles.

[pwolfram]

@mrocklin, For option 1, should the command be dask.set_options(get=dask.async.get_sync)? I'm on 0.8.0

[mrocklin]

Yes, my apologies for the typo.

[shoyer]

I'm pretty sure we now have a thread lock around all writes to NetCDF files, but it's possible that isn't aggressive enough (maybe we can't safely read and write a different file at the same time?). If your script works with synchronous execution I'll take another look.

[pwolfram]

I can't fully confirm that the above scripts works with synchronous execution because the job ran out of its 16hr run time. However, it does appear to be the case that forcing synchronous execution resolves potential issues because previous runs of the script crashed and this one did not. I'll have to try more cases with synchronous execution, especially over the next half week, to see if I encounter more issues but am suspicious this is the problem.

@mrocklin and I noted that the netCDF reader has problems when threading is on when we were using distributed, so this appears to be a likely candidate. We got the same NetCDF: HDF error error as above, and were able to resolve the issue by forcing distributed to work synchronously (non-threaded). @mrocklin should feel free to correct me if I've miss-represented our findings yesterday.

I'm suspicious that the netCDF reader is not thread safe and may not have been compiled as such (http://hdf-forum.184993.n3.nabble.com/Activate-thread-safe-and-enable-cxx-in-HDF5-td2993951.html) but there appear other potential issues that could be part of the problem, e.g., Unidata/netcdf4-python#279 because I am doing so many reads. It may also be possible, as you note @shoyer, that the tread locks aren't aggressive enough.

It would probably be good to come up with some type of testing strategy to better isolate the problem... I'll have to give this more thought.

[mrocklin]

To be clear, we ran into the NetCDF: HDF error error when having multiple threads in the same process open-read-close many different files. I don't think there was any concurrent access of the same file. The problem went away when we switched to using processes rather than threads.

[shoyer]

I did a little digging into this and I'm pretty sure the issue here is that HDF5 cannot do multi-threading -- at all. Moreover, many HDF5 builds are not thread safe.

Right now, we use a single shared lock for all reads with xarray, but for writes we rely on dask.array.store, which only uses different locks for each array it writes. Because @pwolfram's HDF5 file includes multiple variables, each of these gets written with their own thread lock -- which means we end up writing to the same file simultaneously from multiple threads.

So what we could really use here is a lock argument to dask.array.store (like dask.array.from_array) that lets us insist on a using a shared lock when we're writing HDF5 files. Also, we may need to share that same lock between reading and writing data -- I'm not 100% sure. But at the very least we definitely need a lock to stop HDF5 from trying to do multi-threaded writes, whether that's to the same or different files.

[pwolfram]

@thanks @shoyer for looking into this further and for figuring out the cause of the problem. @mrocklin, does this mean that I should submit a dask issue?