fix(chunk-grids): one invariant for zero-length axes across model, clamps, and metadata

changes/4334.bugfix.md

@@ -0,0 +1 @@
+Zero-length dimensions are now handled by a single rule instead of a per-spelling patch: a chunk edge length is always at least 1, while a dimension's extent may be 0 (such a dimension simply has zero chunks). Every way of asking for "one chunk covering the axis" — `chunks=-1`, `chunks=False`, `chunks="auto"`, and `shards="auto"` — now derives the chunk size from the same helper, so they agree on chunk size 1 for a zero-length axis in both Zarr formats. Zarr format 2 metadata now applies the same rule: a stored chunk edge length of 0 on a zero-length axis — which zarr-python 2.x wrote for `chunks=False`, `chunks=-1` and `chunks=(0,)` and then could never read, write, append to or resize (every operation raised `ZeroDivisionError`), and which 3.0–3.3 opened but lost data on append — is read as 1 with a `ZarrUserWarning`, making such arrays usable, while a chunk edge of 0 on an axis that has data is rejected with a clear error. Rectilinear chunk grids (`chunks=[[...], ...]`) can now be created on a zero-length dimension: since no list of positive edge lengths can sum to 0, the given edge lengths are stored as-is and describe the chunks the dimension will grow into on `append` or `resize`, exactly the state a rectilinear dimension is in after being resized down to 0. The private `FixedDimension(size=0, ...)` model, which previously carried its own zero-size special cases, now raises `ValueError`.

docs/user-guide/arrays.md

@@ -708,6 +708,12 @@ z.append(np.arange(10, dtype='float64'))
 print(f"After append:  shape={z.shape}, chunk_sizes={z.write_chunk_sizes}")
 ```
 
+A rectilinear array can also be created with a zero-length dimension: because no
+list of positive chunk sizes can sum to 0, the chunk sizes given for such a
+dimension are stored as-is and describe the chunks the dimension will grow into
+on `append` or `resize` — the same state as resizing an existing rectilinear
+dimension down to 0.
+
 ### Compressors and filters
 
 Rectilinear arrays work with all codecs — compressors, filters, and checksums.

src/zarr/core/chunk_grids.py

@@ -45,22 +45,24 @@
 @dataclass(frozen=True)
 class FixedDimension:
     """Uniform chunk size. Boundary chunks contain less data but are
-    encoded at full size by the codec pipeline."""
+    encoded at full size by the codec pipeline.
 
-    size: int  # chunk edge length (>= 0)
-    extent: int  # array dimension length
+    The chunk edge length is always at least 1, matching the invariant the
+    metadata layer enforces for every stored chunk grid. The extent may be 0:
+    a zero-length axis simply has zero chunks (``ceildiv(0, size) == 0``).
+    """
+
+    size: int  # chunk edge length (>= 1)
+    extent: int  # array dimension length (>= 0)
     nchunks: int = field(init=False, repr=False)
     ngridcells: int = field(init=False, repr=False)
 
     def __post_init__(self) -> None:
-        if self.size < 0:
-            raise ValueError(f"FixedDimension size must be >= 0, got {self.size}")
+        if self.size < 1:
+            raise ValueError(f"FixedDimension size must be >= 1, got {self.size}")
         if self.extent < 0:
             raise ValueError(f"FixedDimension extent must be >= 0, got {self.extent}")
-        if self.size == 0:
-            n = 0
-        else:
-            n = ceildiv(self.extent, self.size)
+        n = ceildiv(self.extent, self.size)
         object.__setattr__(self, "nchunks", n)
         object.__setattr__(self, "ngridcells", n)
 
@@ -69,8 +71,6 @@ def index_to_chunk(self, idx: int) -> int:
             raise IndexError(f"Negative index {idx} is not allowed")
         if idx >= self.extent:
             raise IndexError(f"Index {idx} is out of bounds for extent {self.extent}")
-        if self.size == 0:
-            return 0
         return idx // self.size
 
     def chunk_offset(self, chunk_ix: int) -> int:
@@ -95,8 +95,6 @@ def data_size(self, chunk_ix: int) -> int:
         Does not validate *chunk_ix* — callers must ensure it is in
         ``[0, nchunks)``. Use ``ChunkGrid.__getitem__`` for safe access.
         """
-        if self.size == 0:
-            return 0
         return max(0, min(self.size, self.extent - chunk_ix * self.size))
 
     @property
@@ -110,8 +108,6 @@ def _unique_edge_lengths(self) -> Iterable[int]:
         return (self.size,)
 
     def indices_to_chunks(self, indices: npt.NDArray[np.intp]) -> npt.NDArray[np.intp]:
-        if self.size == 0:
-            return np.zeros_like(indices)
         return indices // self.size
 
     def with_extent(self, new_extent: int) -> FixedDimension:
@@ -640,6 +636,20 @@ class ChunkLayout(NamedTuple):
     inner: ChunkLayout | None = None
 
 
+def _full_span_chunk_size(span: int) -> int:
+    """The edge length of one chunk covering an entire axis of length *span*.
+
+    This is *the* definition of "one chunk spans the axis" for a possibly
+    zero-length axis. Chunk edge lengths must be at least 1 (the invariant
+    shared by `FixedDimension`, `VaryingDimension` and the stored chunk grid
+    metadata), so a zero-length axis gets chunk size 1 and zero chunks. Every
+    spelling that derives a chunk size from a span — ``chunks=-1``,
+    ``chunks=False``, ``chunks="auto"``, ``shards="auto"`` — must route
+    through this helper rather than clamping on its own.
+    """
+    return max(span, 1)
+
+
 def _guess_regular_chunks(
     shape: tuple[int, ...] | int,
     typesize: int,
@@ -677,11 +687,10 @@ def _guess_regular_chunks(
         shape = (shape,)
 
     if typesize == 0:
-        return shape
+        return tuple(_full_span_chunk_size(s) for s in shape)
 
     ndims = len(shape)
-    # require chunks to have non-zero length for all dimensions
-    chunks = np.maximum(np.array(shape, dtype="=f8"), 1)
+    chunks = np.array([_full_span_chunk_size(s) for s in shape], dtype="=f8")
 
     # Determine the optimal chunk size in bytes using a PyTables expression.
     # This is kept as a float.
@@ -724,13 +733,21 @@ def normalize_chunks_1d(chunks: int | Iterable[object], span: int) -> DimensionG
     the span, and the uniform form is O(1) in the number of chunks — a
     dimension with `2**62` chunks must not materialize one entry per chunk.
 
-    `-1` means "one chunk covering the entire span."
+    `-1` means "one chunk covering the entire span" (see `_full_span_chunk_size`
+    for what that means on a zero-length span).
     Explicit chunk size lists must sum to the span exactly and always produce
     `VaryingDimension`, even when the sizes happen to be uniform: the input
     syntax declares the grid kind, so a per-chunk list is preserved as a
     rectilinear dimension rather than silently collapsed to a regular one,
     which would change how the dimension grows on resize. For scalar sizes
     the last chunk may overhang the span.
+
+    The one exception to the sum rule is a zero-length span: no list of
+    positive edges can sum to 0, so any non-empty list is accepted verbatim
+    and the edges describe the chunks the axis will grow into on `append` /
+    `resize`. This is the same state a rectilinear axis reaches when it is
+    resized down to 0 — `VaryingDimension` allows trailing edges beyond the
+    extent — so creating at length 0 and shrinking to 0 are indistinguishable.
     """
     # `numbers.Integral` rather than `int` so that numpy integer scalars (which are not
     # `int` subclasses) take the uniform-chunk path instead of being treated as a sequence.
@@ -741,9 +758,7 @@ def normalize_chunks_1d(chunks: int | Iterable[object], span: int) -> DimensionG
         if chunk_size < -1 or chunk_size == 0:
             raise ValueError(f"Chunk size must be positive or -1, got {chunk_size}")
         if chunk_size == -1:
-            # A zero-length span still gets chunk size 1 (chunk sizes must be positive),
-            # matching the auto-chunking clamp in _guess_regular_chunks.
-            return FixedDimension(size=max(span, 1), extent=span)
+            return FixedDimension(size=_full_span_chunk_size(span), extent=span)
         return FixedDimension(size=chunk_size, extent=span)
     else:
         try:
@@ -768,7 +783,9 @@ def normalize_chunks_1d(chunks: int | Iterable[object], span: int) -> DimensionG
         ints: list[int] = [int(c) for c in chunk_list]  # type: ignore[call-overload]
         if any(c <= 0 for c in ints):
             raise ValueError(f"All chunk sizes must be positive, got {ints}")
-        if sum(ints) != span:
+        # A zero-length span cannot be covered by positive edges; the edges are the
+        # chunks the axis will grow into, exactly as after ``resize(0)``.
+        if span > 0 and sum(ints) != span:
             raise ValueError(f"Chunk sizes {ints} do not sum to span {span}")
         return VaryingDimension(ints, extent=span)
 
@@ -809,7 +826,7 @@ def normalize_chunks_nd(
         )
 
     # handle no chunking: one chunk covering every axis. Routed through the -1 sentinel so
-    # the zero-length-axis clamp lives in one place (normalize_chunks_1d).
+    # the zero-length-axis rule lives in one place (_full_span_chunk_size).
     if chunks is False:
         chunks = -1
 
@@ -864,8 +881,10 @@ def _guess_num_chunks_per_axis_shard(
     In other words the shard would be a (2,2,2) grid of (2,2,2) chunks
     i.e., prod(chunk_shape) * (returned_val ** len(chunk_shape)) * item_size = 256 bytes.
 
-    Degenerate chunk shapes — a 0-dimensional shape, or one containing a zero-length
-    axis — return 1, as the search loop's stopping conditions can never be met.
+    Degenerate inputs — a 0-dimensional chunk shape, or a zero-byte chunk (``item_size``
+    of 0; chunk edge lengths themselves are always at least 1) — return 1, as the
+    search loop's stopping conditions can never be met. A zero-length *array* axis
+    needs no special case: the array-bound check fails immediately for it.
 
     Parameters
     ----------
@@ -886,8 +905,8 @@ def _guess_num_chunks_per_axis_shard(
     if max_bytes < bytes_per_chunk:
         return 1
     num_axes = len(chunk_shape)
-    # For a 0-dimensional chunk shape or one with a zero-length axis, both loop
-    # conditions below are constant, so the loop would never terminate.
+    # For a 0-dimensional chunk shape or a zero-byte chunk, both loop conditions
+    # below are constant, so the loop would never terminate.
     if num_axes == 0 or bytes_per_chunk == 0:
         return 1
     chunks_per_shard = 1

src/zarr/core/metadata/v2.py

@@ -89,6 +89,30 @@ def __init__(
         """
         shape_parsed = parse_shapelike(shape)
         chunks_parsed = parse_shapelike(chunks)
+        # Same invariant as the Zarr format 3 chunk grid metadata: every chunk edge
+        # length is at least 1. zarr-python 2.x wrote `chunks: [0]` for a zero-length
+        # axis created with `chunks=False`, `-1` or `(0,)`, and then could not read,
+        # write, append to or resize the array (every operation divided by zero);
+        # 3.0-3.3 opened such documents but lost data on append. The axis holds no
+        # chunks, so the edge is normalized to 1 — the grid every other "one chunk
+        # spans the axis" spelling produces — which makes the array usable at last.
+        # On an axis that has data, 0 is invalid and any data was never stored.
+        normalized_chunks: list[int] = []
+        for dim_idx, (extent, chunk) in enumerate(zip(shape_parsed, chunks_parsed, strict=False)):
+            if chunk < 1:
+                if chunk < 0 or extent != 0:
+                    raise ValueError(
+                        f"Dimension {dim_idx}: chunk edge length must be >= 1, got {chunk}"
+                    )
+                warnings.warn(
+                    f"Dimension {dim_idx}: chunk edge length 0 on a zero-length axis "
+                    "(as written by zarr-python 2.x) is treated as 1.",
+                    ZarrUserWarning,
+                    stacklevel=2,
+                )
+                chunk = 1
+            normalized_chunks.append(chunk)
+        chunks_parsed = tuple(normalized_chunks) + chunks_parsed[len(shape_parsed) :]
         compressor_parsed = parse_compressor(compressor)
         order_parsed = parse_indexing_order(order)
         dimension_separator_parsed = parse_separator(dimension_separator)