feat: subchunk write order

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.

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

@@ -27,7 +27,7 @@
     Zlib,
     Zstd,
 )
-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
@@ -43,6 +43,7 @@
     "GzipCodec",
     "ShardingCodec",
     "ShardingCodecIndexLocation",
+    "SubchunkWriteOrder",
     "TransposeCodec",
     "VLenBytesCodec",
     "VLenUTF8Codec",

src/zarr/codecs/sharding.py

@@ -5,7 +5,7 @@
 from enum import Enum
 from functools import lru_cache
 from operator import itemgetter
-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
@@ -47,7 +47,6 @@
     BasicIndexer,
     ChunkProjection,
     SelectorTuple,
-    _morton_order,
     _morton_order_keys,
     c_order_iter,
     get_indexer,
@@ -59,7 +58,7 @@
 
 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
@@ -78,6 +77,15 @@ 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)
 
@@ -305,7 +313,9 @@ 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,
@@ -314,16 +324,24 @@ 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
 
@@ -336,14 +354,15 @@ 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))
@@ -523,6 +542,22 @@ async def _decode_partial_single(
         else:
             return out
 
+    def _subchunk_order_iter(self, chunks_per_shard: tuple[int, ...]) -> Iterable[tuple[int, ...]]:
+        match self.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,
@@ -540,8 +575,7 @@ async def _encode_single(
                 chunk_grid=RegularChunkGrid(chunk_shape=chunk_shape),
             )
         )
-
-        shard_builder = dict.fromkeys(morton_order_iter(chunks_per_shard))
+        shard_builder = dict.fromkeys(self._subchunk_order_iter(chunks_per_shard))
 
         await self.codec_pipeline.write(
             [
@@ -582,7 +616,7 @@ 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(np.ndindex(chunks_per_shard))
         else:
             shard_reader = await self._load_full_shard_maybe(
                 byte_getter=byte_setter,
@@ -592,7 +626,7 @@ 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(np.ndindex(chunks_per_shard)))
             )
 
         await self.codec_pipeline.write(
@@ -631,7 +665,7 @@ 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):
             value = map.get(chunk_coords)
             if value is None:
                 continue

tests/test_codecs/test_sharding.py

@@ -1,6 +1,6 @@
 import pickle
 import re
-from typing import Any
+from typing import Any, get_args
 
 import numpy as np
 import numpy.typing as npt
@@ -13,12 +13,15 @@
 from zarr.abc.store import Store
 from zarr.codecs import (
     BloscCodec,
+    BytesCodec,
+    Crc32cCodec,
     ShardingCodec,
     ShardingCodecIndexLocation,
     TransposeCodec,
 )
+from zarr.codecs.sharding import SubchunkWriteOrder, _ShardReader
 from zarr.core.buffer import NDArrayLike, default_buffer_prototype
-from zarr.storage import StorePath, ZipStore
+from zarr.storage import MemoryStore, StorePath, ZipStore
 
 from ..conftest import ArrayRequest
 from .test_codecs import _AsyncArrayProxy, order_from_dim
@@ -555,3 +558,133 @@ def test_sharding_mixed_integer_list_indexing(store: Store) -> None:
     s3 = sharded[0:5, 1, 0:3]
     assert c3.shape == s3.shape == (5, 3)  # type: ignore[union-attr]
     np.testing.assert_array_equal(c3, s3)
+
+
+async def stored_data_and_get_order(
+    codec: ShardingCodec, chunks_per_shard: tuple[int, ...]
+) -> list[tuple[int, ...]]:
+    shard_shape = tuple(c * s for c, s in zip(chunks_per_shard, codec.chunk_shape, strict=True))
+    store = MemoryStore()
+    arr = zarr.create_array(
+        StorePath(store),
+        shape=shard_shape,
+        dtype="uint8",
+        chunks=shard_shape,
+        serializer=codec,
+        filters=None,
+        compressors=None,
+        fill_value=0,
+    )
+
+    arr[:] = np.arange(np.prod(shard_shape), dtype="uint8").reshape(shard_shape)
+
+    shard_buf = await store.get("c/0/0", prototype=default_buffer_prototype())
+    if shard_buf is None:
+        raise RuntimeError("data write failed")
+    index = (await _ShardReader.from_bytes(shard_buf, codec, chunks_per_shard)).index
+    offset_to_coord: dict[int, tuple[int, ...]] = dict(
+        zip(
+            index.get_chunk_slices_vectorized(np.array(list(np.ndindex(chunks_per_shard))))[
+                0
+            ],  # start
+            list(np.ndindex(chunks_per_shard)),  # coord
+            strict=True,
+        )
+    )
+
+    # The physical write order is recovered by sorting coordinates by start offset.
+    return [coord for _, coord in sorted(offset_to_coord.items())]
+
+
+@pytest.mark.parametrize(
+    "subchunk_write_order",
+    get_args(SubchunkWriteOrder),
+)
+async def test_encoded_subchunk_write_order(subchunk_write_order: SubchunkWriteOrder) -> None:
+    """Subchunks must be physically laid out in the shard in the order specified by
+    ``subchunk_write_order``.  We verify this by decoding the shard index and sorting
+    the chunk coordinates by their byte offset."""
+    # Use a non-square chunks_per_shard so all three orderings are distinguishable.
+    chunks_per_shard = (3, 2)
+    chunk_shape = (4, 4)
+    seed = 0
+    codec = ShardingCodec(
+        chunk_shape=chunk_shape,
+        codecs=[BytesCodec()],
+        index_codecs=[BytesCodec(), Crc32cCodec()],
+        index_location=ShardingCodecIndexLocation.end,
+        subchunk_write_order=subchunk_write_order,
+        rng=np.random.default_rng(seed=seed),
+    )
+
+    actual_order = await stored_data_and_get_order(codec, chunks_per_shard)
+    if subchunk_write_order != "unordered":
+        expected_order = list(codec._subchunk_order_iter(chunks_per_shard))
+        assert actual_order == expected_order
+    else:
+        same_order_same_seed = list(
+            ShardingCodec(
+                chunk_shape=chunk_shape,
+                codecs=[BytesCodec()],
+                index_codecs=[BytesCodec(), Crc32cCodec()],
+                index_location=ShardingCodecIndexLocation.end,
+                subchunk_write_order=subchunk_write_order,
+                rng=np.random.default_rng(seed=seed),
+            )._subchunk_order_iter(chunks_per_shard)
+        )
+        assert actual_order == same_order_same_seed
+
+
+async def test_unordered_can_be_seeded() -> None:
+    orders = []
+    chunks_per_shard = (3, 2)
+    chunk_shape = (4, 4)
+    seed = 0
+    for _ in range(4):
+        codec = ShardingCodec(
+            chunk_shape=chunk_shape,
+            codecs=[BytesCodec()],
+            index_codecs=[BytesCodec(), Crc32cCodec()],
+            index_location=ShardingCodecIndexLocation.end,
+            subchunk_write_order="unordered",
+            rng=np.random.default_rng(seed=seed),
+        )
+        # The physical write order is recovered by sorting coordinates by start offset.
+        orders.append(await stored_data_and_get_order(codec, chunks_per_shard))
+    assert all(orders[0] == o for o in orders)
+
+
+@pytest.mark.parametrize(
+    "subchunk_write_order",
+    get_args(SubchunkWriteOrder),
+)
+@pytest.mark.parametrize("do_partial", [True, False], ids=["partial", "complete"])
+def test_subchunk_write_order_roundtrip(
+    subchunk_write_order: SubchunkWriteOrder, do_partial: bool
+) -> None:
+    """Data written with any ``subchunk_write_order`` must round-trip correctly."""
+    chunks_per_shard = (3, 2)
+    chunk_shape = (4, 4)
+    shard_shape = tuple(c * s for c, s in zip(chunks_per_shard, chunk_shape, strict=True))
+    data = np.arange(np.prod(shard_shape), dtype="uint16").reshape(shard_shape)
+    arr = zarr.create_array(
+        StorePath(MemoryStore()),
+        shape=shard_shape,
+        dtype=data.dtype,
+        chunks=shard_shape,
+        serializer=ShardingCodec(
+            chunk_shape=chunk_shape,
+            codecs=[BytesCodec()],
+            subchunk_write_order=subchunk_write_order,
+        ),
+        filters=None,
+        compressors=None,
+        fill_value=0,
+    )
+    if do_partial:
+        sub_data = data[: (shard_shape[0] // 2)]
+        arr[: (shard_shape[0] // 2)] = data[: (shard_shape[0] // 2)]
+        data = np.vstack([sub_data, np.zeros_like(sub_data)])
+    else:
+        arr[:] = data
+    np.testing.assert_array_equal(arr[:], data)