diff --git a/xrspatial/geotiff/_writer.py b/xrspatial/geotiff/_writer.py
index ab0bd069d..a4e2a757e 100644
--- a/xrspatial/geotiff/_writer.py
+++ b/xrspatial/geotiff/_writer.py
@@ -396,6 +396,24 @@ def _write(data: np.ndarray, path: str, *,
         entry_point="_write",
     )
 
+    # Issue #2311: defense in depth for the COG auto-overview hang.
+    # ``to_geotiff`` already rejects non-positive tile_size when either
+    # tiled or cog is true, but ``_write`` is a public-ish array-level
+    # entry point reached from the GPU CPU-fallback path and downstream
+    # code that imports it directly. Without the gate, a non-positive
+    # tile_size with ``cog=True`` hits ``ZeroDivisionError`` in
+    # ``_write_tiled`` (tiled=True path) or the infinite auto-overview
+    # halving loop (tiled=False path). Convert both to a typed
+    # ValueError up front so callers see one actionable failure mode
+    # instead of two divergent ones.
+    if cog and (not isinstance(tile_size, (int, np.integer))
+                or isinstance(tile_size, bool)
+                or tile_size <= 0):
+        raise ValueError(
+            f"tile_size must be a positive int for cog=True (consumed by "
+            f"tile encoding and auto-overview generation), got "
+            f"tile_size={tile_size!r}.")
+
     # Issue #2138 gap #7: auto-promote ``float16`` and ``bool_`` before
     # the dtype mapper. ``to_geotiff`` already does this upstream; the
     # push-down here lets direct callers feed unsupported dtypes without
@@ -496,10 +514,27 @@ def _write(data: np.ndarray, path: str, *,
             # holds actual decimation factors (2, 4, 8, ...) so the loop
             # below treats auto-generated and user-supplied lists
             # identically (issue #1766).
+            #
+            # Defense in depth (issue #2311): require a positive tile_size
+            # before entering the halving loop. The public ``to_geotiff``
+            # entry point already validates tile_size when either tiled or
+            # cog is true, but ``write()`` is also reachable from internal
+            # callers; without this guard a non-positive tile_size leaves
+            # the ``oh > tile_size`` termination condition permanently true
+            # while the inner ``oh > 0`` guard suppresses appends, so the
+            # loop spins forever.
+            if (not isinstance(tile_size, (int, np.integer))
+                    or isinstance(tile_size, bool)
+                    or tile_size <= 0):
+                raise ValueError(
+                    f"tile_size must be a positive int for COG overview "
+                    f"generation, got tile_size={tile_size!r}.")
             overview_levels = []
             oh, ow = h, w
             factor = 2
-            while oh > tile_size and ow > tile_size and len(overview_levels) < _MAX_OVERVIEW_LEVELS:
+            while (oh > tile_size and ow > tile_size
+                    and oh > 0 and ow > 0
+                    and len(overview_levels) < _MAX_OVERVIEW_LEVELS):
                 oh //= 2
                 ow //= 2
                 if oh > 0 and ow > 0:
diff --git a/xrspatial/geotiff/_writers/eager.py b/xrspatial/geotiff/_writers/eager.py
index acb8e6b43..3cc6b9247 100644
--- a/xrspatial/geotiff/_writers/eager.py
+++ b/xrspatial/geotiff/_writers/eager.py
@@ -335,10 +335,16 @@ def to_geotiff(data: xr.DataArray | np.ndarray,
         raise TypeError(
             f"nodata must be numeric (int or float), got {nodata!r}")
 
-    # tiled=False ignores tile_size, so only validate when tiled output
-    # is requested. Shared with write_geotiff_gpu via
+    # tiled=False ignores tile_size for the strip-layout pixel data, so
+    # historically validation only ran when tiled=True. The COG path
+    # (cog=True) still reads tile_size to auto-generate overviews in
+    # ``_writer.py`` regardless of the strip-vs-tiled choice, so a
+    # non-positive tile_size with cog=True drove the overview loop
+    # into a hang once oh, ow halved to 0 (issue #2311). Validate
+    # tile_size whenever either path will consume it: tiled output OR
+    # COG overview generation. Shared with write_geotiff_gpu via
     # _validate_tile_size_arg so both writers keep identical validation.
-    if tiled:
+    if tiled or cog:
         _validate_tile_size_arg(tile_size)
 
     _validate_nodata_arg(nodata)
diff --git a/xrspatial/geotiff/tests/test_cog_tile_size_hang_2311.py b/xrspatial/geotiff/tests/test_cog_tile_size_hang_2311.py
new file mode 100644
index 000000000..60056cd2b
--- /dev/null
+++ b/xrspatial/geotiff/tests/test_cog_tile_size_hang_2311.py
@@ -0,0 +1,210 @@
+"""COG writer rejects non-positive ``tile_size`` regardless of ``tiled`` (#2311).
+
+Before this fix, ``to_geotiff(..., cog=True, tiled=False, tile_size=<=0)``
+hung the writer. ``tile_size`` validation only ran when ``tiled=True``, but
+the COG path in ``_writer.py`` still used ``tile_size`` to auto-generate
+overviews regardless of ``tiled``. With ``tile_size=-1`` the auto-overview
+loop in ``_writer.py:490`` had ``oh > tile_size and ow > tile_size``
+permanently true once ``oh, ow`` halved to 0, while the inner
+``if oh > 0 and ow > 0`` guard prevented the level list from growing --
+the loop never exited.
+
+The fix lives in two places:
+
+1. ``to_geotiff`` in ``_writers/eager.py`` now runs ``_validate_tile_size_arg``
+   when ``tiled=True`` OR ``cog=True``. The COG path consumes ``tile_size``
+   for overview generation regardless of strip-vs-tiled layout, so the
+   public boundary must validate it in both cases.
+2. The auto-overview loop in ``_writer.py`` has a defensive pre-check that
+   raises if ``tile_size`` is not a positive int, plus a tightened loop
+   condition that requires ``oh, ow > 0`` to continue. Together these mean
+   the loop cannot run forever even if a future internal caller bypasses
+   the public validator.
+
+Each row below uses a SIGALRM-based timeout so a regression that brings
+the hang back fails the test instead of locking up the run. SIGALRM is a
+POSIX-only mechanism (CPython on Linux/macOS); the tests fall back to
+plain execution on Windows, where the original hang is still a concern
+but the watchdog is unavailable.
+"""
+from __future__ import annotations
+
+import contextlib
+import os
+import signal
+import warnings
+
+import numpy as np
+import pytest
+import xarray as xr
+
+from xrspatial.geotiff import to_geotiff
+
+
+@contextlib.contextmanager
+def _alarm_timeout(seconds: int):
+    """Raise TimeoutError after ``seconds`` to bound test failure modes.
+
+    No-op on platforms that lack SIGALRM (Windows). The window is large
+    enough that a healthy raise path finishes well before the alarm
+    fires; if the fix regresses the writer hangs and the alarm fires.
+    """
+    if not hasattr(signal, 'SIGALRM') or os.name == 'nt':
+        yield
+        return
+
+    def _handler(signum, frame):  # noqa: ARG001
+        raise TimeoutError(
+            f'test exceeded {seconds}s watchdog; the writer likely '
+            f'regressed into the #2311 infinite-loop hang.'
+        )
+
+    old = signal.signal(signal.SIGALRM, _handler)
+    signal.alarm(seconds)
+    try:
+        yield
+    finally:
+        signal.alarm(0)
+        signal.signal(signal.SIGALRM, old)
+
+
+def _float_da(shape=(64, 64)):
+    """A small float32 DataArray large enough to trigger COG overview build."""
+    return xr.DataArray(
+        np.zeros(shape, dtype=np.float32), dims=('y', 'x')
+    )
+
+
+# ---------------------------------------------------------------------------
+# Public boundary: ``to_geotiff(..., cog=True, tile_size<=0)`` must raise.
+# Covers both tiled=True and tiled=False, plus 0 and a negative value, so
+# the validator gate stays on regardless of layout flag.
+# ---------------------------------------------------------------------------
+
+@pytest.mark.parametrize('tiled', [True, False])
+@pytest.mark.parametrize('tile_size', [-1, 0])
+def test_to_geotiff_cog_non_positive_tile_size_raises(tmp_path, tiled, tile_size):
+    """``cog=True`` with ``tile_size<=0`` raises ValueError up front,
+    regardless of ``tiled``. Before #2311 this hung the writer when
+    ``tiled=False``."""
+    da = _float_da()
+    p = tmp_path / f'cog_tile_size_hang_2311_t{int(tiled)}_ts{tile_size}.tif'
+
+    with _alarm_timeout(5), pytest.raises(ValueError) as exc:
+        to_geotiff(da, str(p), cog=True, tiled=tiled, tile_size=tile_size)
+
+    msg = str(exc.value)
+    assert 'tile_size' in msg, msg
+    # The shared validator says "positive int" -- pin the substring so a
+    # message rewrite still keeps the actionable wording.
+    assert 'positive' in msg.lower(), msg
+
+
+# ---------------------------------------------------------------------------
+# Sanity: ``cog=False`` with ``tiled=False`` still accepts an unused
+# ``tile_size`` (the existing "ignored" warning shape) -- the new gate
+# must not fire when neither path will consume the value.
+# ---------------------------------------------------------------------------
+
+def test_to_geotiff_non_cog_strip_does_not_validate_tile_size(tmp_path):
+    """When neither tiled output nor COG overview generation will use
+    ``tile_size``, the validator gate stays off. The pre-existing
+    "tile_size ignored" warning still fires (it carries its own
+    non-default-value check, not a positivity check), but no error
+    is raised."""
+    da = _float_da()
+    p = tmp_path / 'cog_tile_size_hang_2311_no_cog_strip.tif'
+
+    # A negative tile_size with cog=False AND tiled=False is accepted
+    # (with the "ignored" warning) because nothing consumes the value.
+    # Use ``filterwarnings`` to swallow the warning so the test only
+    # asserts no raise / no hang.
+    with _alarm_timeout(5), warnings.catch_warnings():
+        warnings.simplefilter('ignore')
+        to_geotiff(da, str(p), cog=False, tiled=False, tile_size=-1)
+
+    assert p.exists(), 'writer should have produced a strip-layout file'
+
+
+# ---------------------------------------------------------------------------
+# Defense in depth: drive the inner writer directly with a bad tile_size
+# and assert the auto-overview loop raises instead of hanging. Guards
+# against future internal callers that bypass ``to_geotiff``'s public
+# validator.
+# ---------------------------------------------------------------------------
+
+@pytest.mark.parametrize('tile_size', [-1, 0])
+def test_writer_auto_overview_loop_rejects_non_positive_tile_size(
+        tmp_path, tile_size):
+    """``_write(..., cog=True, overview_levels=None)`` raises ValueError
+    when ``tile_size`` is not a positive int, instead of spinning in the
+    halving loop. The public ``to_geotiff`` already validates earlier;
+    this is the inner-writer safety net (#2311)."""
+    from xrspatial.geotiff._writer import _write
+
+    # Minimal float32 array large enough for the auto-overview branch to
+    # be entered. The exact pixel values do not matter -- the validator
+    # check runs before any encoding work.
+    data = np.zeros((64, 64), dtype=np.float32)
+    out = tmp_path / f'cog_tile_size_hang_2311_inner_ts{tile_size}.tif'
+
+    with _alarm_timeout(5), pytest.raises(ValueError) as exc:
+        _write(data, str(out),
+               compression='none',
+               tiled=True,
+               tile_size=tile_size,
+               cog=True,
+               overview_levels=None)
+
+    assert 'tile_size' in str(exc.value), str(exc.value)
+
+
+# ---------------------------------------------------------------------------
+# Non-int tile_size values reach the same gate. The public
+# ``_validate_tile_size`` (called from ``to_geotiff`` when tiled or cog is
+# true) rejects None, float, and bool with typed errors; the
+# defense-in-depth gate at the top of ``_write`` does the same for direct
+# callers. Both layers should reject all three types.
+# ---------------------------------------------------------------------------
+
+@pytest.mark.parametrize('bad_tile_size', [None, 128.0, True, False])
+def test_to_geotiff_cog_non_int_tile_size_raises(tmp_path, bad_tile_size):
+    """Non-int ``tile_size`` (None, float, bool) with ``cog=True`` is
+    rejected at the public boundary, regardless of ``tiled``. Bool is
+    explicitly listed because Python treats ``True``/``False`` as int
+    subclasses (#2311 follow-up)."""
+    da = _float_da()
+    p = tmp_path / (
+        f'cog_tile_size_hang_2311_nonint_{type(bad_tile_size).__name__}.tif')
+
+    with _alarm_timeout(5), pytest.raises((ValueError, TypeError)) as exc:
+        to_geotiff(da, str(p), cog=True, tiled=True, tile_size=bad_tile_size)
+
+    assert 'tile_size' in str(exc.value), str(exc.value)
+
+
+# ---------------------------------------------------------------------------
+# Inner-loop guard coverage: confirm the auto-overview halving loop's own
+# ``tile_size > 0`` pre-check is present in ``_write``'s compiled
+# constants. Inspecting the constants pins the literal so a future
+# refactor that removes the inner guard fails this test loudly even if
+# the top-of-``_write`` gate still catches the bad input at runtime.
+# (Reaching the inner guard through ``_write`` directly would require
+# patching out the top gate, which is invasive; the constants check is
+# the simplest reliable pin without rewriting production code.)
+# ---------------------------------------------------------------------------
+
+def test_inner_overview_loop_guard_message_is_pinned():
+    """Pin the inner-overview ``tile_size`` guard literal so removing
+    the loop-side defense fails this test even when the top gate at
+    line 407 still raises for the same inputs (#2311)."""
+    from xrspatial.geotiff import _writer as wmod
+
+    guard_msg = (
+        'tile_size must be a positive int for COG overview '
+        'generation, got tile_size=')
+    consts = wmod._write.__code__.co_consts
+    found = any(isinstance(c, str) and guard_msg in c for c in consts)
+    assert found, (
+        'inner-loop guard message not present in _write constants; the '
+        'auto-overview guard introduced in #2311 may have been removed.')