From d864330d832ea02b53828fd82f792fc82d68cb17 Mon Sep 17 00:00:00 2001
From: Fix 11417 <fix-11417@users.noreply.github.com>
Date: Fri, 26 Jun 2026 22:46:31 +0800
Subject: [PATCH] fix: skip CF encoding for zarr arrays with native DateTime64
 dtype (GH#11350)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

When writing to a zarr array with native DateTime64 dtype (zarr v3),
skip CF datetime encoding. CF encoding converts datetime64 to int64
with units/calendar attributes, which corrupts data when written to
a native DateTime64 zarr array — the int64 values get misinterpreted
as raw epoch seconds instead of the encoded 'days since' values.

Fixes #11350
---
 xarray/backends/zarr.py       | 25 +++++++++++++++++++++++--
 xarray/tests/test_backends.py | 29 +++++++++++++++++++++++++++++
 2 files changed, 52 insertions(+), 2 deletions(-)

diff --git a/xarray/backends/zarr.py b/xarray/backends/zarr.py
index 6f3e1ad4eb4..5b34b6be2d6 100644
--- a/xarray/backends/zarr.py
+++ b/xarray/backends/zarr.py
@@ -1090,8 +1090,29 @@ def store(
                 )
                 vars_with_encoding[vn] = variables[vn].copy(deep=False)
                 vars_with_encoding[vn].encoding = existing_vars[vn].encoding
-            vars_with_encoding, _ = self.encode(vars_with_encoding, {})
-            variables_encoded.update(vars_with_encoding)
+
+            # Skip CF encoding for zarr v3 native DateTime64 arrays (GH#11350).
+            native_datetime_vars: set[str] = set()
+            if _zarr_v3() and self.zarr_group.metadata.zarr_format == 3:
+                from zarr.dtype import DateTime64
+
+                for vn in existing_variable_names:
+                    name = _encode_variable_name(vn)
+                    zarr_array = self.members[name]
+                    if isinstance(zarr_array.metadata.data_type, DateTime64):
+                        native_datetime_vars.add(vn)
+
+            vars_to_encode = {
+                vn: v
+                for vn, v in vars_with_encoding.items()
+                if vn not in native_datetime_vars
+            }
+            if vars_to_encode:
+                encoded, _ = self.encode(vars_to_encode, {})
+                variables_encoded.update(encoded)
+
+            for vn in native_datetime_vars:
+                variables_encoded[vn] = vars_with_encoding[vn]
 
             for var_name in existing_variable_names:
                 variables_encoded[var_name] = _validate_and_transpose_existing_dims(
diff --git a/xarray/tests/test_backends.py b/xarray/tests/test_backends.py
index 4e08b71260b..1bc3e143c02 100644
--- a/xarray/tests/test_backends.py
+++ b/xarray/tests/test_backends.py
@@ -8153,3 +8153,32 @@ def test_get_mtime_non_file_paths() -> None:
 
     # GDAL virtual filesystem paths are not real files
     assert _get_mtime("/vsicurl/https://example.com/file.nc") is None
+
+
+@requires_zarr
+@requires_zarr_v3
+def test_zarr_datetime64_roundtrip_no_corruption() -> None:
+    """Regression test for https://github.com/pydata/xarray/issues/11350.
+
+    Writing a dataset back to a zarr store with native DateTime64 dtype
+    should not corrupt the datetime values.
+    """
+    import zarr.dtype
+
+    store = zarr.storage.MemoryStore()
+    g = zarr.create_group(store)
+    a = g.create_array(
+        "a",
+        shape=(2,),
+        dtype=zarr.dtype.DateTime64(unit="s", scale_factor=1),
+        dimension_names=["time"],
+    )
+    a[:] = np.array(["2025-01-01", "2025-01-02"], dtype="<M8[us]")
+
+    ds = xr.open_zarr(store, chunks=None, consolidated=False).compute()
+    expected = ds["a"].values.copy()
+
+    ds.to_zarr(store, mode="r+", consolidated=False)
+
+    ds_roundtripped = xr.open_zarr(store, chunks=None, consolidated=False).compute()
+    np.testing.assert_array_equal(ds_roundtripped["a"].values, expected)