Skip to content

A reactive notebook for Python — run reproducible experiments, query with SQL, execute as a script, deploy as an app, and version with git. Stored as pure Python. All in a modern, AI-native editor.

fix(mpl): disconnect toolbar callbacks on cell rerun (#9531)

Kiran Gadhavefb38a2521e86
diff --git a/marimo/_plugins/ui/_impl/from_mpl_interactive.py b/marimo/_plugins/ui/_impl/from_mpl_interactive.py
index d4fcb63106..bf6bd3c7b0 100644
--- a/marimo/_plugins/ui/_impl/from_mpl_interactive.py
+++ b/marimo/_plugins/ui/_impl/from_mpl_interactive.py
@@ -29,6 +29,31 @@ if TYPE_CHECKING:
LOGGER = _loggers.marimo_logger()
+
+def _disconnect_owner_callbacks(canvas: Any, owner: Any) -> None:
+ """Disconnect every callback on ``canvas.callbacks`` whose receiver is ``owner``.
+
+ ``FigureCanvasBase.callbacks`` is a property returning
+ ``figure._canvas_callbacks``, so every canvas bound to the same
+ figure shares one registry. ``NavigationToolbar2.__init__``
+ registers ``_zoom_pan_handler`` (and a few siblings) on this
+ shared registry; matplotlib has no public counterpart that
+ disconnects them when a canvas/toolbar is discarded. Without this
+ cleanup, every cell rerun would stack another toolbar's handler on
+ the registry, leading to duplicate ``press_pan``/``release_pan``
+ dispatch.
+ """
+ registry = canvas.callbacks
+ cids: list[int] = []
+ for handlers in registry.callbacks.values():
+ for cid, ref in handlers.items():
+ fn = ref() if callable(ref) else ref
+ if getattr(fn, "__self__", None) is owner:
+ cids.append(cid)
+ for cid in cids:
+ registry.disconnect(cid)
+
+
# Must match the className on the container div in MplInteractivePlugin.tsx
_MPL_SCOPE = ".mpl-interactive-figure"
@@ -328,27 +353,33 @@ class _MplCleanupHandle(CellLifecycleItem):
def dispose(self, context: RuntimeContext, deletion: bool) -> bool:
del context, deletion
- # Disconnect the web socket from the figure manager first,
- # then destroy the canvas to prevent stale timer callbacks
- # on thread teardown (avoids "QObject::killTimer" errors).
- if self._figure_manager is not None and self._sync_ws is not None:
- try:
- self._figure_manager.remove_web_socket(self._sync_ws)
- except Exception:
- pass
- if self._figure_manager is not None:
+ fm = self._figure_manager
+ if fm is not None:
+ if self._sync_ws is not None:
+ try:
+ fm.remove_web_socket(self._sync_ws)
+ except Exception:
+ LOGGER.exception("Failed to remove mpl web socket")
+
+ canvas = fm.canvas
+ toolbar = getattr(canvas, "toolbar", None)
+ if toolbar is not None:
+ try:
+ _disconnect_owner_callbacks(canvas, toolbar)
+ except Exception:
+ LOGGER.exception(
+ "Failed to disconnect mpl toolbar callbacks"
+ )
+
try:
# get the root figure (in case of Subfigure) which handles dpi
- root = self._figure_manager.canvas.figure.figure
+ root = canvas.figure.figure
root.set_dpi(self._original_dpi)
root.set_size_inches(*self._original_size_inches)
except Exception:
- pass
-
- try:
- self._figure_manager.canvas.close()
- except Exception:
- pass
+ LOGGER.exception(
+ "Failed to restore mpl figure dpi/size on dispose"
+ )
self._comm.close()
return True
diff --git a/tests/_plugins/ui/_impl/test_from_mpl_interactive.py b/tests/_plugins/ui/_impl/test_from_mpl_interactive.py
index a475e15d57..6533eff0db 100644
--- a/tests/_plugins/ui/_impl/test_from_mpl_interactive.py
+++ b/tests/_plugins/ui/_impl/test_from_mpl_interactive.py
@@ -281,6 +281,9 @@ class TestMplCleanupHandle:
mock_comm = MagicMock()
mock_manager = MagicMock()
+ # The disconnect helper iterates this; keep it empty so the mock
+ # doesn't trip over MagicMock's default value generators.
+ mock_manager.canvas.callbacks.callbacks = {}
mock_ws = MagicMock()
handle = _MplCleanupHandle(
mock_comm,
@@ -293,7 +296,6 @@ class TestMplCleanupHandle:
assert result is True
mock_manager.remove_web_socket.assert_called_once_with(mock_ws)
- mock_manager.canvas.close.assert_called_once()
mock_comm.close.assert_called_once()
def test_dispose_tolerates_manager_errors(self) -> None:
@@ -304,7 +306,10 @@ class TestMplCleanupHandle:
mock_comm = MagicMock()
mock_manager = MagicMock()
mock_manager.remove_web_socket.side_effect = RuntimeError("boom")
- mock_manager.canvas.close.side_effect = RuntimeError("boom")
+ # Force the disconnect path to blow up too.
+ type(mock_manager.canvas).callbacks = property(
+ lambda _self: (_ for _ in ()).throw(RuntimeError("boom"))
+ )
mock_ws = MagicMock()
handle = _MplCleanupHandle(
mock_comm,
@@ -320,6 +325,73 @@ class TestMplCleanupHandle:
mock_comm.close.assert_called_once()
[email protected]("matplotlib")
+class TestToolbarCallbackCleanup:
+ """Disposing a cleanup handle must disconnect the toolbar's callbacks
+ from the figure-shared event registry.
+
+ `canvas.callbacks` is a property that delegates to
+ `figure._canvas_callbacks`, so every canvas bound to the same figure
+ shares one registry. Without explicit cleanup, each cell rerun stacks
+ another `_zoom_pan_handler` on the registry — leading to duplicate
+ `press_pan`/`release_pan` dispatch and ultimately the
+ `AttributeError: 'Axes' object has no attribute '_pan_start'` from
+ matplotlib's `Axes.end_pan`.
+ """
+
+ @staticmethod
+ def _count_toolbar_handlers(fig: Any, signal: str) -> int:
+ handlers = fig.canvas.callbacks.callbacks.get(signal, {})
+ count = 0
+ for ref in handlers.values():
+ fn = ref() if callable(ref) else ref
+ owner = getattr(fn, "__self__", None)
+ if (
+ owner is not None
+ and type(owner).__name__ == "NavigationToolbar2WebAgg"
+ ):
+ count += 1
+ return count
+
+ def test_dispose_disconnects_toolbar_callbacks(self) -> None:
+ import matplotlib.pyplot as plt
+
+ from marimo._plugins.ui._impl.from_mpl_interactive import (
+ _MplCleanupHandle,
+ mpl_interactive,
+ )
+
+ fig, ax = plt.subplots()
+ ax.plot([1, 2, 3])
+
+ with patch("marimo._plugins.ui._impl.comm.broadcast_notification"):
+ elem1 = mpl_interactive(fig)
+ assert self._count_toolbar_handlers(fig, "button_press_event") == 1
+ assert self._count_toolbar_handlers(fig, "button_release_event") == 1
+
+ # Simulate cell teardown: marimo's runtime invokes dispose on the
+ # lifecycle handle when the cell re-runs or is deleted.
+ cleanup = _MplCleanupHandle(
+ comm=elem1._comm,
+ figure_manager=elem1._figure_manager,
+ sync_ws=elem1._sync_ws,
+ original_dpi=elem1._original_dpi,
+ original_size_inches=elem1._original_size_inches,
+ )
+ cleanup.dispose(context=MagicMock(), deletion=False)
+
+ # Simulate the cell re-running: a new mpl_interactive on the same
+ # figure object. After dispose, only the new toolbar's callbacks
+ # should be live on the shared registry.
+ with patch("marimo._plugins.ui._impl.comm.broadcast_notification"):
+ _elem2 = mpl_interactive(fig)
+
+ assert self._count_toolbar_handlers(fig, "button_press_event") == 1
+ assert self._count_toolbar_handlers(fig, "button_release_event") == 1
+
+ plt.close(fig)
+
+
@pytest.mark.requires("matplotlib")
class TestMplInteractiveArgs:
"""Test that mpl_interactive passes correct args to the UIElement."""