(fix): use typesize on Blosc codec

changes/2962.fix.rst

@@ -0,0 +1 @@
+Internally use `typesize` constructor parameter for :class:`numcodecs.blosc.Blosc` to improve compression ratios back to the v2-package levels.

docs/user-guide/arrays.rst

@@ -209,8 +209,8 @@ prints additional diagnostics, e.g.::
    Serializer         : BytesCodec(endian=<Endian.little: 'little'>)
    Compressors        : (BloscCodec(typesize=4, cname=<BloscCname.zstd: 'zstd'>, clevel=3, shuffle=<BloscShuffle.bitshuffle: 'bitshuffle'>, blocksize=0),)
    No. bytes          : 400000000 (381.5M)
-   No. bytes stored   : 9696520
-   Storage ratio      : 41.3
+   No. bytes stored   : 3558573
+   Storage ratio      : 112.4
    Chunks Initialized : 100
 
 .. note::

src/zarr/codecs/blosc.py

@@ -8,6 +8,7 @@
 
 import numcodecs
 from numcodecs.blosc import Blosc
+from packaging.version import Version
 
 from zarr.abc.codec import BytesBytesCodec
 from zarr.core.buffer.cpu import as_numpy_array_wrapper
@@ -163,6 +164,9 @@ def _blosc_codec(self) -> Blosc:
             "shuffle": map_shuffle_str_to_int[self.shuffle],
             "blocksize": self.blocksize,
         }
+        # See https://github.com/zarr-developers/numcodecs/pull/713
+        if Version(numcodecs.__version__) >= Version("0.16.0"):
+            config_dict["typesize"] = self.typesize
         return Blosc.from_config(config_dict)
 
     async def _decode_single(

tests/test_codecs/test_blosc.py

@@ -1,7 +1,10 @@
 import json
+import platform
 
+import numcodecs
 import numpy as np
 import pytest
+from packaging.version import Version
 
 import zarr
 from zarr.abc.store import Store
@@ -54,3 +57,29 @@ async def test_blosc_evolve(store: Store, dtype: str) -> None:
         assert blosc_configuration_json["shuffle"] == "bitshuffle"
     else:
         assert blosc_configuration_json["shuffle"] == "shuffle"
+
+
+async def test_typesize() -> None:
+    a = np.arange(1000000)
+    codecs = [zarr.codecs.BytesCodec(), zarr.codecs.BloscCodec()]
+    z = zarr.array(a, chunks=(10000), codecs=codecs)
+    bytes = (await z.store.get("c/0", prototype=default_buffer_prototype())).to_bytes()
+    size = len(bytes)
+    msg = f"Blosc size mismatch.  First 10 bytes: {bytes[:20]} and last 10 bytes: {bytes[-20:]}"
+    match (
+        Version(numcodecs.__version__) >= Version("0.16.0"),
+        platform.system() == "Windows",
+        Version(np.__version__) < Version("2.0.0"),
+    ):
+        case True, True, True:
+            # See https://github.com/zarr-developers/zarr-python/pull/2962
+            # for why this condition is distinct.  It's not clear
+            # if it's the python version of the numpy version with windows.
+            expected_size = 400
+            assert size == 400, msg
+        case (True, True, False) | (True, False, _):
+            expected_size = 402
+            assert size == expected_size, msg
+        case False, _, _:
+            expected_size = 10216
+            assert size == expected_size, msg