LRUStoreCache

changes/3357.feature.rst

@@ -0,0 +1,3 @@
+Add LRUStoreCache for improved performance with remote stores
+
+The new ``LRUStoreCache`` provides a least-recently-used (LRU) caching layer that can be wrapped around any zarr store to significantly improve performance, especially for remote stores where network latency is a bottleneck.

docs/user-guide/lrustorecache.rst

@@ -0,0 +1,186 @@
+.. only:: doctest
+
+   >>> import shutil
+   >>> shutil.rmtree('test.zarr', ignore_errors=True)
+
+.. _user-guide-lrustorecache:
+
+LRUStoreCache guide
+===================
+
+The :class:`zarr.storage.LRUStoreCache` provides a least-recently-used (LRU) cache layer
+that can be wrapped around any Zarr store to improve performance for repeated data access.
+This is particularly useful when working with remote stores (e.g., S3, HTTP) where network
+latency can significantly impact data access speed.
+
+The LRUStoreCache implements a cache that stores frequently accessed data chunks in memory,
+automatically evicting the least recently used items when the cache reaches its maximum size.
+
+.. note::
+   The LRUStoreCache is a wrapper store that maintains compatibility with the full
+   :class:`zarr.abc.store.Store` API while adding transparent caching functionality.
+
+Basic Usage
+-----------
+
+Creating an LRUStoreCache is straightforward - simply wrap any existing store with the cache:
+
+   >>> import zarr
+   >>> import zarr.storage
+   >>> import numpy as np
+   >>>
+   >>> # Create a local store and wrap it with LRU cache
+   >>> local_store = zarr.storage.LocalStore('test.zarr')
+   >>> cache = zarr.storage.LRUStoreCache(local_store, max_size=1024 * 1024 * 256)  # 256MB cache
+   >>>
+   >>> # Create an array using the cached store
+   >>> zarr_array = zarr.zeros((100, 100), chunks=(10, 10), dtype='f8', store=cache, mode='w')
+   >>>
+   >>> # Write some data to force chunk creation
+   >>> zarr_array[:] = np.random.random((100, 100))
+
+The ``max_size`` parameter controls the maximum memory usage of the cache in bytes. Set it to
+``None`` for unlimited cache size (use with caution).
+
+Performance Benefits
+--------------------
+
+The LRUStoreCache provides significant performance improvements for repeated data access:
+
+   >>> import time
+   >>>
+   >>> # Benchmark reading with cache
+   >>> start = time.time()
+   >>> for _ in range(100):
+   ...     _ = zarr_array[:]
+   >>> elapsed_cache = time.time() - start
+   >>>
+   >>> # Compare with direct store access (without cache)
+   >>> zarr_array_nocache = zarr.open('test.zarr', mode='r')
+   >>> start = time.time()
+   >>> for _ in range(100):
+   ...     _ = zarr_array_nocache[:]
+   >>> elapsed_nocache = time.time() - start
+   >>>
+   >>> speedup = elapsed_nocache/elapsed_cache
+
+Cache effectiveness is particularly pronounced with repeated access to the same data chunks.
+
+Remote Store Caching
+--------------------
+
+The LRUStoreCache is most beneficial when used with remote stores where network latency
+is a significant factor. Here's a conceptual example::
+
+   # Example with a remote store (requires gcsfs)
+   import gcsfs
+
+   # Create a remote store (Google Cloud Storage example)
+   gcs = gcsfs.GCSFileSystem(token='anon')
+   remote_store = gcsfs.GCSMap(
+       root='your-bucket/data.zarr',
+       gcs=gcs,
+       check=False
+   )
+
+   # Wrap with LRU cache for better performance
+   cached_store = zarr.storage.LRUStoreCache(remote_store, max_size=2**28)
+
+   # Open array through cached store
+   z = zarr.open(cached_store)
+
+The first access to any chunk will be slow (network retrieval), but subsequent accesses
+to the same chunk will be served from the local cache, providing dramatic speedup.
+
+Cache Configuration
+-------------------
+
+The LRUStoreCache can be configured with several parameters:
+
+**max_size**: Controls the maximum memory usage of the cache in bytes
+
+   >>> # Create a base store for demonstration
+   >>> store = zarr.storage.LocalStore('config_example.zarr')
+   >>>
+   >>> # 256MB cache
+   >>> cache = zarr.storage.LRUStoreCache(store, max_size=2**28)
+   >>>
+   >>> # Unlimited cache size (use with caution)
+   >>> cache = zarr.storage.LRUStoreCache(store, max_size=None)
+
+**read_only**: Create a read-only cache
+
+   >>> cache = zarr.storage.LRUStoreCache(store, max_size=2**28, read_only=True)
+
+Cache Statistics
+----------------
+
+The LRUStoreCache provides statistics to monitor cache performance:
+
+   >>> # Access some data to generate cache activity
+   >>> data = zarr_array[0:50, 0:50]  # First access - cache miss
+   >>> data = zarr_array[0:50, 0:50]  # Second access - cache hit
+   >>>
+   >>> cache_hits = cache.hits
+   >>> cache_misses = cache.misses
+   >>> total_requests = cache.hits + cache.misses
+   >>> cache_hit_ratio = cache.hits / total_requests if total_requests > 0 else 0
+   >>> # Typical hit ratio is > 50% with repeated access patterns
+
+Cache Management
+----------------
+
+The cache provides methods for manual cache management:
+
+   >>> # Clear all cached values but keep keys cache
+   >>> cache.invalidate_values()
+   >>>
+   >>> # Clear keys cache
+   >>> cache.invalidate_keys()
+   >>>
+   >>> # Clear entire cache
+   >>> cache.invalidate()
+
+Best Practices
+--------------
+
+1. **Size the cache appropriately**: Set ``max_size`` based on available memory and expected data access patterns
+2. **Use with remote stores**: The cache provides the most benefit when wrapping slow remote stores
+3. **Monitor cache statistics**: Use hit/miss ratios to tune cache size and access patterns
+4. **Consider data locality**: Access data in chunks sequentially rather than jumping around randomly to maximize cache reuse
+
+Examples from Real Usage
+------------------------
+
+Here's a complete example demonstrating cache effectiveness:
+
+   >>> import zarr
+   >>> import zarr.storage
+   >>> import time
+   >>> import numpy as np
+   >>>
+   >>> # Create test data
+   >>> local_store = zarr.storage.LocalStore('benchmark.zarr')
+   >>> cache = zarr.storage.LRUStoreCache(local_store, max_size=2**28)
+   >>> zarr_array = zarr.zeros((100, 100), chunks=(10, 10), dtype='f8', store=cache, mode='w')
+   >>> zarr_array[:] = np.random.random((100, 100))
+   >>>
+   >>> # Demonstrate cache effectiveness with repeated access
+   >>> # First access (cache miss):
+   >>> start = time.time()
+   >>> data = zarr_array[20:30, 20:30]
+   >>> first_access = time.time() - start
+   >>>
+   >>> # Second access (cache hit):
+   >>> start = time.time()
+   >>> data = zarr_array[20:30, 20:30]  # Same data should be cached
+   >>> second_access = time.time() - start
+   >>>
+   >>> # Calculate cache performance metrics
+   >>> cache_speedup = first_access/second_access
+
+This example shows how the LRUStoreCache can significantly reduce access times for repeated
+data reads, particularly important when working with remote data sources.
+
+.. _Zip Store Specification: https://github.com/zarr-developers/zarr-specs/pull/311
+.. _fsspec: https://filesystem-spec.readthedocs.io

src/zarr/storage/__init__.py

@@ -4,6 +4,7 @@
 from typing import Any
 
 from zarr.errors import ZarrDeprecationWarning
+from zarr.storage._cache import LRUStoreCache
 from zarr.storage._common import StoreLike, StorePath
 from zarr.storage._fsspec import FsspecStore
 from zarr.storage._local import LocalStore
@@ -16,6 +17,7 @@
 __all__ = [
     "FsspecStore",
     "GpuMemoryStore",
+    "LRUStoreCache",
     "LocalStore",
     "LoggingStore",
     "MemoryStore",