Regression in 3.3.0: numpy integer chunk sizes raise TypeError: 'numpy.int64' object is not iterable #4255

Description

Since 3.3.0, passing a chunk size as a numpy integer raises a TypeError. This worked in 3.2.1 and earlier.

The cause is a disagreement between two functions in src/zarr/core/chunk_grids.py about what counts as an integer:

• normalize_chunks_nd (line 790) tests isinstance(chunks, numbers.Integral)
• normalize_chunks_1d (line 734) tests isinstance(chunks, int)

np.int64 registers as numbers.Integral but is not a subclass of int. So a per-dimension numpy integer passes the outer dispatch, reaches normalize_chunks_1d, fails the isinstance(chunks, int) narrowing, and falls into the else branch intended for explicit per-dimension chunk sequences — where list(chunks) is called on a scalar.

Both lines are still present on main (bd3e398).

Reproduction

import numpy as np
import zarr

zarr.zeros((4, 4), dtype="f4", chunks=(np.int64(2), np.int64(2)))

File "zarr/core/chunk_grids.py", line 742, in normalize_chunks_1d
    chunk_list = list(chunks)
                 ^^^^^^^^^^^^
TypeError: 'numpy.int64' object is not iterable

Mixed tuples fail the same way, and only on the numpy-typed entries — (1, 3, np.int64(16), np.int64(16)) raises, (1, 3, 16, 16) does not.

Version info

• zarr 3.3.0 (fails); 3.2.1 (works)
• numpy 2.5.2
• Python 3.12, Linux

Introduced by #3899 (refactor: simplify internal chunk representation), which landed in 3.3.0.

Why this matters in practice

Numpy integers arise naturally whenever a chunk shape is computed rather than written as a literal, since any numpy reduction or elementwise op yields numpy scalars. For example, a downstream library sizing chunks from the array shape:

plane_shape = np.minimum(shape[-2:], 32768)   # numpy array
chunks = (1, depth, *plane_shape)             # trailing entries are np.int64

This is what we hit: ultrack's large_chunk_size builds its chunk tuple this way and passes it to zarr.zeros, so every array-creation call through it broke on 3.3.0. We are currently pinned to zarr<3.3 as a workaround.

Because the failure is a TypeError about iteration deep inside chunk-grid code, it is quite hard to trace back to "your chunk sizes are numpy integers" — the error names neither chunks nor the offending value.

Suggested fix

Widen the narrowing in normalize_chunks_1d to match its caller:

-    if isinstance(chunks, int):
+    if isinstance(chunks, numbers.Integral):
+        chunks = int(chunks)

This is consistent with the rest of the module: guess_chunks already ends with return tuple(int(x) for x in chunks), and the else branch already accepts numbers.Integral elements and coerces them with int(c). Only the scalar path is strict.

Happy to open a PR with the fix plus a regression test covering numpy scalar chunk sizes, if that would be useful.