Merge branch 'main' into indexing-test-cleanup

Author: d-v-b

changes/2929.bugfix.md

@@ -0,0 +1,3 @@
+Fix equality comparison of `ArrayV2Metadata` and `ArrayV3Metadata` objects with a
+`NaN` fill value. Such objects are now compared by their JSON-serialized form, so two
+otherwise-identical metadata objects with a `NaN` (or infinite) fill value compare equal.

changes/3826.feature.md

@@ -0,0 +1 @@
+Added a `subchunk_write_order` option to `ShardingCodec` to allow for `morton`, `unordered`, `lexicographic`, and `colexicographic` subchunk orderings.

changes/3973.removal.md

@@ -0,0 +1 @@
+Removed the NumPy 1.x implementation of the `VariableLengthUTF8` data type because NumPy 1.x is no longer supported under [SPEC0](https://scientific-python.org/specs/spec-0000/).

docs/user-guide/performance.md

@@ -113,6 +113,13 @@ bytes within chunks of an array may improve the compression ratio, depending on
 the structure of the data, the compression algorithm used, and which compression
 filters (e.g., byte-shuffle) have been applied.
 
+### Subchunk memory layout
+
+The order of chunks **within each shard** can be changed via the `subchunk_write_order` parameter of the `ShardingCodec`. That parameter is a string which must be one of `["morton", "lexicographic", "colexicographic", "unordered"]`.
+
+By default [`morton`](https://en.wikipedia.org/wiki/Z-order_curve) order provides good spatial locality however [`lexicographic` (i.e., row-major)](https://en.wikipedia.org/wiki/Row-_and_column-major_order), for example, may be better suited to "batched" workflows where some form of sequential reading through a fixed number of outer dimensions is desired. The options are `lexicographic`, `morton`, `unordered` (i.e., random), and `colexicographic`.
+
+
 ### Empty chunks
 
 It is possible to configure how Zarr handles the storage of chunks that are "empty"

src/zarr/codecs/__init__.py

@@ -29,7 +29,7 @@
     Zstd,
 )
 from zarr.codecs.scale_offset import ScaleOffset
-from zarr.codecs.sharding import ShardingCodec, ShardingCodecIndexLocation
+from zarr.codecs.sharding import ShardingCodec, ShardingCodecIndexLocation, SubchunkWriteOrder
 from zarr.codecs.transpose import TransposeCodec
 from zarr.codecs.vlen_utf8 import VLenBytesCodec, VLenUTF8Codec
 from zarr.codecs.zstd import ZstdCodec
@@ -47,6 +47,7 @@
     "ScaleOffset",
     "ShardingCodec",
     "ShardingCodecIndexLocation",
+    "SubchunkWriteOrder",
     "TransposeCodec",
     "VLenBytesCodec",
     "VLenUTF8Codec",

src/zarr/codecs/sharding.py

@@ -4,7 +4,7 @@
 from dataclasses import dataclass, replace
 from enum import Enum
 from functools import lru_cache
-from typing import TYPE_CHECKING, Any, NamedTuple, cast
+from typing import TYPE_CHECKING, Any, Literal, NamedTuple, cast
 
 import numpy as np
 import numpy.typing as npt
@@ -46,8 +46,6 @@
     BasicIndexer,
     ChunkProjection,
     SelectorTuple,
-    _morton_order,
-    _morton_order_keys,
     c_order_iter,
     get_indexer,
     morton_order_iter,
@@ -64,7 +62,7 @@
 
 if TYPE_CHECKING:
     from collections.abc import Iterator
-    from typing import Self
+    from typing import Final, Self
 
     from zarr.core.common import JSON
     from zarr.core.dtype.wrapper import TBaseDType, TBaseScalar, ZDType
@@ -83,6 +81,15 @@ class ShardingCodecIndexLocation(Enum):
     end = "end"
 
 
+SubchunkWriteOrder = Literal["morton", "unordered", "lexicographic", "colexicographic"]
+SUBCHUNK_WRITE_ORDER: Final[tuple[str, str, str, str]] = (
+    "morton",
+    "unordered",
+    "lexicographic",
+    "colexicographic",
+)
+
+
 def parse_index_location(data: object) -> ShardingCodecIndexLocation:
     return parse_enum(data, ShardingCodecIndexLocation)
 
@@ -272,14 +279,13 @@ def to_dict_vectorized(
         dict mapping chunk coordinate tuples to Buffer or None
         """
         starts, ends, valid = self.index.get_chunk_slices_vectorized(chunk_coords_array)
-        chunk_coords_keys = _morton_order_keys(self.index.chunks_per_shard)
 
         result: dict[tuple[int, ...], Buffer | None] = {}
-        for i, coords in enumerate(chunk_coords_keys):
+        for i, coords in enumerate(chunk_coords_array):
             if valid[i]:
-                result[coords] = self.buf[int(starts[i]) : int(ends[i])]
+                result[tuple(coords.ravel())] = self.buf[int(starts[i]) : int(ends[i])]
             else:
-                result[coords] = None
+                result[tuple(coords.ravel())] = None
 
         return result
 
@@ -293,7 +299,9 @@ class ShardingCodec(
     chunk_shape: tuple[int, ...]
     codecs: tuple[Codec, ...]
     index_codecs: tuple[Codec, ...]
+    rng: np.random.Generator | None
     index_location: ShardingCodecIndexLocation = ShardingCodecIndexLocation.end
+    subchunk_write_order: SubchunkWriteOrder = "morton"
 
     def __init__(
         self,
@@ -302,16 +310,24 @@ def __init__(
         codecs: Iterable[Codec | dict[str, JSON]] = (BytesCodec(),),
         index_codecs: Iterable[Codec | dict[str, JSON]] = (BytesCodec(), Crc32cCodec()),
         index_location: ShardingCodecIndexLocation | str = ShardingCodecIndexLocation.end,
+        subchunk_write_order: SubchunkWriteOrder = "morton",
+        rng: np.random.Generator | None = None,
     ) -> None:
         chunk_shape_parsed = parse_shapelike(chunk_shape)
         codecs_parsed = parse_codecs(codecs)
         index_codecs_parsed = parse_codecs(index_codecs)
         index_location_parsed = parse_index_location(index_location)
+        if subchunk_write_order not in SUBCHUNK_WRITE_ORDER:
+            raise ValueError(
+                f"Unrecognized subchunk write order: {subchunk_write_order}. Only {SUBCHUNK_WRITE_ORDER} are allowed."
+            )
 
         object.__setattr__(self, "chunk_shape", chunk_shape_parsed)
         object.__setattr__(self, "codecs", codecs_parsed)
         object.__setattr__(self, "index_codecs", index_codecs_parsed)
         object.__setattr__(self, "index_location", index_location_parsed)
+        object.__setattr__(self, "subchunk_write_order", subchunk_write_order)
+        object.__setattr__(self, "rng", rng)
 
         # Use instance-local lru_cache to avoid memory leaks
 
@@ -324,14 +340,15 @@ def __init__(
 
     # todo: typedict return type
     def __getstate__(self) -> dict[str, Any]:
-        return self.to_dict()
+        return {"rng": self.rng, **self.to_dict()}
 
     def __setstate__(self, state: dict[str, Any]) -> None:
         config = state["configuration"]
         object.__setattr__(self, "chunk_shape", parse_shapelike(config["chunk_shape"]))
         object.__setattr__(self, "codecs", parse_codecs(config["codecs"]))
         object.__setattr__(self, "index_codecs", parse_codecs(config["index_codecs"]))
         object.__setattr__(self, "index_location", parse_index_location(config["index_location"]))
+        object.__setattr__(self, "rng", state["rng"])
 
         # Use instance-local lru_cache to avoid memory leaks
         # object.__setattr__(self, "_get_chunk_spec", lru_cache()(self._get_chunk_spec))
@@ -509,6 +526,24 @@ async def _decode_partial_single(
         else:
             return out
 
+    def _subchunk_order_iter(
+        self, chunks_per_shard: tuple[int, ...], subchunk_write_order: SubchunkWriteOrder
+    ) -> Iterable[tuple[int, ...]]:
+        match subchunk_write_order:
+            case "morton":
+                subchunk_iter = morton_order_iter(chunks_per_shard)
+            case "lexicographic":
+                subchunk_iter = np.ndindex(chunks_per_shard)
+            case "colexicographic":
+                subchunk_iter = (c[::-1] for c in np.ndindex(chunks_per_shard[::-1]))
+            case "unordered":
+                subchunk_list = list(np.ndindex(chunks_per_shard))
+                (self.rng if self.rng is not None else np.random.default_rng()).shuffle(
+                    subchunk_list
+                )
+                subchunk_iter = iter(subchunk_list)
+        return subchunk_iter
+
     async def _encode_single(
         self,
         shard_array: NDBuffer,
@@ -526,8 +561,7 @@ async def _encode_single(
                 chunk_grid=ChunkGrid.from_sizes(shard_shape, chunk_shape),
             )
         )
-
-        shard_builder = dict.fromkeys(morton_order_iter(chunks_per_shard))
+        shard_builder = dict.fromkeys(self._subchunk_order_iter(chunks_per_shard, "lexicographic"))
 
         await self.codec_pipeline.write(
             [
@@ -570,7 +604,7 @@ async def _encode_partial_single(
         )
 
         if self._is_complete_shard_write(indexer, chunks_per_shard):
-            shard_dict = dict.fromkeys(morton_order_iter(chunks_per_shard))
+            shard_dict = dict.fromkeys(self._subchunk_order_iter(chunks_per_shard, "lexicographic"))
         else:
             shard_reader = await self._load_full_shard_maybe(
                 byte_getter=byte_setter,
@@ -580,7 +614,7 @@ async def _encode_partial_single(
             shard_reader = shard_reader or _ShardReader.create_empty(chunks_per_shard)
             # Use vectorized lookup for better performance
             shard_dict = shard_reader.to_dict_vectorized(
-                np.asarray(_morton_order(chunks_per_shard))
+                np.array(list(self._subchunk_order_iter(chunks_per_shard, "lexicographic")))
             )
 
         await self.codec_pipeline.write(
@@ -619,7 +653,7 @@ async def _encode_shard_dict(
 
         template = buffer_prototype.buffer.create_zero_length()
         chunk_start = 0
-        for chunk_coords in morton_order_iter(chunks_per_shard):
+        for chunk_coords in self._subchunk_order_iter(chunks_per_shard, self.subchunk_write_order):
             value = map.get(chunk_coords)
             if value is None:
                 continue