Serialize numpy scalar attributes as their Python equivalents

Setting an attribute to a numpy scalar such as np.float32(1.5), np.int64(3)
or np.bool_(True) failed with "Object of type float32 is not JSON
serializable". np.float64 only worked because it subclasses float.

json_to_buffer now passes a default hook to json.dumps that converts
numpy scalars with .item(). Other unsupported types still raise TypeError.
Add a round-trip test covering both zarr formats. Fixes #3478.

Author: vijay-loopr

changes/3478.bugfix.md

@@ -0,0 +1 @@
+Attributes holding numpy scalars, such as `np.float32(1.5)`, `np.int64(3)` or `np.bool_(True)`, can be written again. They were previously rejected with `Object of type float32 is not JSON serializable`.

src/zarr/core/_json.py

@@ -25,7 +25,9 @@
 from __future__ import annotations
 
 import json
-from typing import TYPE_CHECKING, cast
+from typing import TYPE_CHECKING, Any, cast
+
+import numpy as np
 
 from zarr.core.buffer import default_buffer_prototype
 
@@ -35,6 +37,13 @@
     from zarr.core.common import JSON
 
 
+def _json_default(obj: Any) -> Any:
+    """Convert a numpy scalar, which `json` cannot encode, to the equivalent Python value."""
+    if isinstance(obj, np.generic):
+        return obj.item()
+    raise TypeError(f"Object of type {type(obj).__name__} is not JSON serializable")
+
+
 def buffer_to_json(buffer: Buffer) -> JSON:
     """Parse the contents of a `Buffer` as a JSON value."""
     # json.loads is typed as returning Any; the result is by definition JSON.
@@ -96,7 +105,9 @@ def json_to_buffer(
     """
     if prototype is None:
         prototype = default_buffer_prototype()
-    return prototype.buffer.from_bytes(json.dumps(obj, indent=indent, allow_nan=allow_nan).encode())
+    return prototype.buffer.from_bytes(
+        json.dumps(obj, indent=indent, allow_nan=allow_nan, default=_json_default).encode()
+    )
 
 
 async def get_json(store: Store, key: str, *, byte_range: ByteRequest | None = None) -> JSON | None:

tests/test_attributes.py

@@ -26,6 +26,20 @@ def test_put(data: dict[str, Any], zarr_format: ZarrFormat) -> None:
     assert deep_nan_equal(dict(attrs), expected)
 
 
+@pytest.mark.parametrize("zarr_format", [2, 3])
+@pytest.mark.parametrize(
+    "value",
+    [np.float32(1.5), np.float32(np.nan), np.int64(3), np.uint8(7), np.bool_(True)],
+)
+def test_numpy_scalar_attribute_roundtrip(value: Any, zarr_format: ZarrFormat) -> None:
+    """numpy scalars are stored as their Python equivalents and survive reopening the store."""
+    store = zarr.storage.MemoryStore()
+    zarr.create_group(store, zarr_format=zarr_format, attributes={"foo": value})
+    reopened = zarr.open_group(store, mode="r", zarr_format=zarr_format)
+    expected = value.item()
+    assert deep_nan_equal({"foo": reopened.attrs["foo"]}, {"foo": expected})
+
+
 def test_asdict() -> None:
     store = zarr.storage.MemoryStore()
     attrs = zarr.core.attributes.Attributes(