feat: subchunk write order

Author: ilan-gold

src/zarr/codecs/sharding.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import random
 from collections.abc import Iterable, Mapping, MutableMapping
 from dataclasses import dataclass, replace
 from enum import Enum
@@ -46,7 +47,6 @@
 from zarr.core.indexing import (
     BasicIndexer,
     SelectorTuple,
-    _morton_order,
     _morton_order_keys,
     c_order_iter,
     get_indexer,
@@ -77,10 +77,27 @@ class ShardingCodecIndexLocation(Enum):
     end = "end"
 
 
+class SubchunkWriteOrder(Enum):
+    """
+    Enum for the order of the chunks within a shard.
+
+    unordered is implemented via `random.shuffle` over the lexicographic order.
+    """
+
+    morton = "morton"
+    unordered = "unordered"
+    lexicographic = "lexicographic"
+    colexicographic = "colexicographic"
+
+
 def parse_index_location(data: object) -> ShardingCodecIndexLocation:
     return parse_enum(data, ShardingCodecIndexLocation)
 
 
+def parse_subchunk_write_order(data: object) -> SubchunkWriteOrder:
+    return parse_enum(data, SubchunkWriteOrder)
+
+
 @dataclass(frozen=True)
 class _ShardingByteGetter(ByteGetter):
     shard_dict: ShardMapping
@@ -305,6 +322,7 @@ class ShardingCodec(
     codecs: tuple[Codec, ...]
     index_codecs: tuple[Codec, ...]
     index_location: ShardingCodecIndexLocation = ShardingCodecIndexLocation.end
+    subchunk_write_order: SubchunkWriteOrder = SubchunkWriteOrder.morton
 
     def __init__(
         self,
@@ -313,16 +331,19 @@ 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 | str = SubchunkWriteOrder.morton,
     ) -> 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)
+        subchunk_write_order_parsed = parse_subchunk_write_order(subchunk_write_order)
 
         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_parsed)
 
         # Use instance-local lru_cache to avoid memory leaks
 
@@ -522,6 +543,20 @@ async def _decode_partial_single(
         else:
             return out
 
+    def _subchunk_iter(self, chunks_per_shard: tuple[int, ...]) -> Iterable[tuple[int, ...]]:
+        match self.subchunk_write_order:
+            case SubchunkWriteOrder.morton:
+                subchunk_iter = morton_order_iter(chunks_per_shard)
+            case SubchunkWriteOrder.lexicographic:
+                subchunk_iter = np.ndindex(chunks_per_shard)
+            case SubchunkWriteOrder.colexicographic:
+                subchunk_iter = (c[::-1] for c in np.ndindex(chunks_per_shard[::-1]))
+            case SubchunkWriteOrder.unordered:
+                subchunk_list = list(np.ndindex(chunks_per_shard))
+                random.shuffle(subchunk_list)
+                subchunk_iter = iter(subchunk_list)
+        return subchunk_iter
+
     async def _encode_single(
         self,
         shard_array: NDBuffer,
@@ -539,8 +574,7 @@ 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_iter(chunks_per_shard))
 
         await self.codec_pipeline.write(
             [
@@ -581,7 +615,9 @@ 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)))
+        shard_dict = shard_reader.to_dict_vectorized(
+            np.asarray(list(self._subchunk_iter(chunks_per_shard)))
+        )
 
         indexer = list(
             get_indexer(
@@ -625,7 +661,7 @@ 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_iter(chunks_per_shard):
             value = map.get(chunk_coords)
             if value is None:
                 continue

tests/test_codecs/test_sharding.py

@@ -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 +558,90 @@ 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)
+
+
+@pytest.mark.parametrize(
+    "subchunk_write_order",
+    list(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)
+    shard_shape = tuple(c * s for c, s in zip(chunks_per_shard, chunk_shape, strict=True))
+
+    codec = ShardingCodec(
+        chunk_shape=chunk_shape,
+        codecs=[BytesCodec()],
+        index_codecs=[BytesCodec(), Crc32cCodec()],
+        index_location=ShardingCodecIndexLocation.end,
+        subchunk_write_order=subchunk_write_order,
+    )
+    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.
+    actual_order = [coord for _, coord in sorted(offset_to_coord.items())]
+    expected_order = list(codec._subchunk_iter(chunks_per_shard))
+    assert (actual_order == expected_order) == (
+        subchunk_write_order != SubchunkWriteOrder.unordered
+    )
+
+
+@pytest.mark.parametrize(
+    "subchunk_write_order",
+    list(SubchunkWriteOrder),
+)
+def test_subchunk_write_order_roundtrip(subchunk_write_order: SubchunkWriteOrder) -> 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,
+    )
+    arr[:] = data
+    np.testing.assert_array_equal(arr[:], data)