xarray.backends refactor (pydata#2261)

* WIP: xarray.backends.file_manager for managing file objects.

This is intended to replace both PickleByReconstructionWrapper and
DataStorePickleMixin with something more compartmentalized.

xref GH2121

* Switch rasterio to use FileManager

* lint fixes

* WIP: rewrite FileManager to always use an LRUCache

* Test coverage

* Don't use move_to_end

* minor clarification

* Switch FileManager.acquire() to a method

* Python 2 compat

* Update xarray.set_options() to add file_cache_maxsize and validation

* Add assert for FILE_CACHE.maxsize

* More docstring for FileManager

* Add accidentally omited tests for LRUCache

* Adapt scipy backend to use FileManager

* Stickler fix

* Fix failure on Python 2.7

* Finish adjusting backends to use FileManager

* Fix bad import

* WIP on distributed

* More WIP

* Fix distributed write tests

* Fixes

* Minor fixup

* whats new

* More refactoring: remove state from backends entirely

* Cleanup

* Fix failing in-memory datastore tests

* Fix inaccessible datastore

* fix autoclose warnings

* Fix PyNIO failures

* No longer disable HDF5 file locking

We longer need to explicitly HDF5_USE_FILE_LOCKING='FALSE' because we
properly close open files.

* whats new and default file cache size

* Whats new tweak

* Refactor default lock logic to backend classes

* Rename get_resource_lock -> get_write_lock

* Don't acquire unnecessary locks in __getitem__

* Fix bad merge

* Fix import

* Remove unreachable code

asv_bench/asv.conf.json

@@ -64,6 +64,7 @@
         "scipy": [""],
         "bottleneck": ["", null],
         "dask": [""],
+        "distributed": [""],
     },
 
 

asv_bench/benchmarks/dataset_io.py

@@ -1,5 +1,7 @@
 from __future__ import absolute_import, division, print_function
 
+import os
+
 import numpy as np
 import pandas as pd
 
@@ -14,6 +16,9 @@
     pass
 
 
+os.environ['HDF5_USE_FILE_LOCKING'] = 'FALSE'
+
+
 class IOSingleNetCDF(object):
     """
     A few examples that benchmark reading/writing a single netCDF file with
@@ -405,3 +410,39 @@ def time_open_dataset_scipy_with_time_chunks(self):
         with dask.set_options(get=dask.multiprocessing.get):
             xr.open_mfdataset(self.filenames_list, engine='scipy',
                               chunks=self.time_chunks)
+
+
+def create_delayed_write():
+    import dask.array as da
+    vals = da.random.random(300, chunks=(1,))
+    ds = xr.Dataset({'vals': (['a'], vals)})
+    return ds.to_netcdf('file.nc', engine='netcdf4', compute=False)
+
+
+class IOWriteNetCDFDask(object):
+    timeout = 60
+    repeat = 1
+    number = 5
+
+    def setup(self):
+        requires_dask()
+        self.write = create_delayed_write()
+
+    def time_write(self):
+        self.write.compute()
+
+
+class IOWriteNetCDFDaskDistributed(object):
+    def setup(self):
+        try:
+            import distributed
+        except ImportError:
+            raise NotImplementedError
+        self.client = distributed.Client()
+        self.write = create_delayed_write()
+
+    def cleanup(self):
+        self.client.shutdown()
+
+    def time_write(self):
+        self.write.compute()

doc/api.rst

@@ -624,3 +624,6 @@ arguments for the ``from_store`` and ``dump_to_store`` Dataset methods:
    backends.H5NetCDFStore
    backends.PydapDataStore
    backends.ScipyDataStore
+   backends.FileManager
+   backends.CachingFileManager
+   backends.DummyFileManager

doc/whats-new.rst

@@ -33,14 +33,27 @@ v0.11.0 (unreleased)
 Breaking changes
 ~~~~~~~~~~~~~~~~
 
+- Xarray's storage backends now automatically open and close files when
+  necessary, rather than requiring opening a file with ``autoclose=True``. A
+  global least-recently-used cache is used to store open files; the default
+  limit of 128 open files should suffice in most cases, but can be adjusted if
+  necessary with
+  ``xarray.set_options(file_cache_maxsize=...)``. The ``autoclose`` argument
+  to ``open_dataset`` and related functions has been deprecated and is now a
+  no-op.
+
+  This change, along with an internal refactor of xarray's storage backends,
+  should significantly improve performance when reading and writing
+  netCDF files with Dask, especially when working with many files or using
+  Dask Distributed. By `Stephan Hoyer <https://github.com/shoyer>`_
+
+Documentation
+~~~~~~~~~~~~~
 - Reduction of :py:meth:`DataArray.groupby` and :py:meth:`DataArray.resample`
   without dimension argument will change in the next release.
   Now we warn a FutureWarning.
   By `Keisuke Fujii <https://github.com/fujiisoup>`_.
 
-Documentation
-~~~~~~~~~~~~~
-
 Enhancements
 ~~~~~~~~~~~~
 

xarray/backends/__init__.py

@@ -4,6 +4,7 @@
 formats. They should not be used directly, but rather through Dataset objects.
 """
 from .common import AbstractDataStore
+from .file_manager import FileManager, CachingFileManager, DummyFileManager
 from .memory import InMemoryDataStore
 from .netCDF4_ import NetCDF4DataStore
 from .pydap_ import PydapDataStore
@@ -15,6 +16,9 @@
 
 __all__ = [
     'AbstractDataStore',
+    'FileManager',
+    'CachingFileManager',
+    'DummyFileManager',
     'InMemoryDataStore',
     'NetCDF4DataStore',
     'PydapDataStore',