LRU store cache

docs/api/storage.rst

@@ -21,6 +21,18 @@ Storage (``zarr.storage``)
     .. automethod:: close
     .. automethod:: flush
 
+.. autoclass:: LRUStoreCache
+
+    .. automethod:: invalidate
+    .. automethod:: invalidate_values
+    .. automethod:: invalidate_keys
+
 .. autofunction:: init_array
 .. autofunction:: init_group
+.. autofunction:: contains_array
+.. autofunction:: contains_group
+.. autofunction:: listdir
+.. autofunction:: rmdir
+.. autofunction:: getsize
+.. autofunction:: rename
 .. autofunction:: migrate_1to2

docs/release.rst

@@ -127,6 +127,11 @@ Enhancements
 * **Added support for ``datetime64`` and ``timedelta64`` data types**;
   :issue:`85`, :issue:`215`.
 
+* **New LRUStoreCache class**. The class :class:`zarr.storage.LRUStoreCache` has been
+  added and provides a means to locally cache data in memory from a store that may be
+  slow, e.g., a store that retrieves data from a remote server via the network;
+  :issue:`223`.
+
 * **New copy functions**. The new functions :func:`zarr.convenience.copy` and
   :func:`zarr.convenience.copy_all` provide a way to copy groups and/or arrays
   between HDF5 and Zarr, or between two Zarr groups. The

docs/tutorial.rst

@@ -729,6 +729,9 @@ group (requires `lmdb <http://lmdb.readthedocs.io/>`_ to be installed)::
     >>> z[:] = 42
     >>> store.close()
 
+Distributed/cloud storage
+~~~~~~~~~~~~~~~~~~~~~~~~~
+
 It is also possible to use distributed storage systems. The Dask project has
 implementations of the ``MutableMapping`` interface for Amazon S3 (`S3Map
 <http://s3fs.readthedocs.io/en/latest/api.html#s3fs.mapping.S3Map>`_), Hadoop
@@ -767,6 +770,37 @@ Here is an example using S3Map to read an array created previously::
     >>> z[:].tostring()
     b'Hello from the cloud!'
 
+Note that retrieving data from a remote service via the network can be significantly
+slower than retrieving data from a local file system, and will depend on network latency
+and bandwidth between the client and server systems. If you are experiencing poor
+performance, there are several things you can try. One option is to increase the array
+chunk size, which will reduce the number of chunks and thus reduce the number of network
+round-trips required to retrieve data for an array (and thus reduce the impact of network
+latency). Another option is to try to increase the compression ratio by changing
+compression options or trying a different compressor (which will reduce the impact of
+limited network bandwidth). As of version 2.2, Zarr also provides the
+:class:`zarr.storage.LRUStoreCache` which can be used to implement a local in-memory cache
+layer over a remote store. E.g.::
+
+    >>> s3 = s3fs.S3FileSystem(anon=True, client_kwargs=dict(region_name='eu-west-2'))
+    >>> store = s3fs.S3Map(root='zarr-demo/store', s3=s3, check=False)
+    >>> cache = zarr.LRUStoreCache(store, max_size=2**28)
+    >>> root = zarr.group(store=cache)
+    >>> z = root['foo/bar/baz']
+    >>> from timeit import timeit
+    >>> # first data access is relatively slow, retrieved from store
+    ... timeit('print(z[:].tostring())', number=1, globals=globals())  # doctest: +SKIP
+    b'Hello from the cloud!'
+    0.1081731989979744
+    >>> # second data access is faster, uses cache
+    ... timeit('print(z[:].tostring())', number=1, globals=globals())  # doctest: +SKIP
+    b'Hello from the cloud!'
+    0.0009490990014455747
+
+If you are still experiencing poor performance with distributed/cloud storage, please
+raise an issue on the GitHub issue tracker with any profiling data you can provide, as
+there may be opportunities to optimise further either within Zarr or within the mapping
+interface to the storage.
 
 .. _tutorial_copy:
 

zarr/__init__.py

@@ -7,7 +7,7 @@
 from zarr.creation import (empty, zeros, ones, full, array, empty_like, zeros_like,
                            ones_like, full_like, open_array, open_like, create)
 from zarr.storage import (DictStore, DirectoryStore, ZipStore, TempStore,
-                          NestedDirectoryStore, DBMStore, LMDBStore)
+                          NestedDirectoryStore, DBMStore, LMDBStore, LRUStoreCache)
 from zarr.hierarchy import group, open_group, Group
 from zarr.sync import ThreadSynchronizer, ProcessSynchronizer
 from zarr.codecs import *

zarr/compat.py

@@ -16,9 +16,16 @@
     class PermissionError(Exception):
         pass
 
+    def OrderedDict_move_to_end(od, key):
+        od[key] = od.pop(key)
+
+
 else:  # pragma: py2 no cover
 
     text_type = str
     binary_type = bytes
     from functools import reduce
     PermissionError = PermissionError
+
+    def OrderedDict_move_to_end(od, key):
+        od.move_to_end(key)