Repository navigation
SWR matplotlib__matplotlib-18454 - #75
richeshgupta wants to merge 1 commit into
Conversation
When converting from jQuery to vanilla JavaScript, the trigger for the figure closure was set to a non-existent event name, as the 'remove' event is jQuery-specific. As there is no DOM node removal event (unless using MutationObserver), connect this handler to the notebook's `Cell` `OutputArea`'s 'cleared' event. This is triggered both by re-running a cell, and by using the Clear Output menu option. Fixes matplotlib#18447.
| mpl.figure.prototype.handle_close = function (fig, msg) { | ||
| var width = fig.canvas.width / fig.ratio; | ||
| fig.root.removeEventListener('remove', this._remove_fig_handler); | ||
| fig.cell_info[0].output_area.element.off( |
There was a problem hiding this comment.
[MEDIUM] Closing one figure unbinds the 'cleared' cleanup handler of every other figure in the same cell
Issue
Every figure shares the single prototype function _remove_fig_handler, and jQuery's .off(events, handler) removes all bindings of that function's guid on the element, so handle_close strips the 'cleared' handler of every figure bound to the cell's output area, not just the figure being closed. After one sibling figure is closed, clearing or re-running the cell never closes the remaining figures' comms, leaving their websockets and kernel-side figure managers alive for the life of the page.
Evidence
CONFIRMED: jQuery's .one() stores a wrapper carrying the original function's guid ('use same guid so caller can remove using origFn') and .off(type, fn) removes every handler with that guid; _remove_fig_handler is one prototype-level function that every figure in a cell binds to the same output_area.element, so a single handle_close removes them all. Creating two figures in one cell, closing the first, then clearing the cell shows the second figure's comm never closing.
| console.error('Failed to find cell for figure', id, fig); | ||
| return; | ||
| } | ||
| fig.cell_info[0].output_area.element.one( |
There was a problem hiding this comment.
[MEDIUM] No test exercises the new nbagg figure teardown when notebook output is cleared
Issue
The PR's core behaviour (closing figures when a cell's output area is cleared, and unbinding that handler in handle_close) is reachable only in a browser notebook session, and the diff adds no test covering it. A wrong event name, a broken event.data.fig contract in _remove_fig_handler, or the unconditional fig.cell_info[0] dereference in handle_close would fail no build, and the silent failure is the figure and websocket leak this teardown exists to prevent.
Evidence
SUSPECTED: the diff touches no test file, but coverage elsewhere in the repo was not verifiable from this review; a browser-driven nbagg test that renders a figure, clears the output area, and asserts the comm/websocket closes, plus one that closes a figure whose cell was never found, would settle it.
SWRBench case matplotlib__matplotlib-18454: pinned base ec45ec7 to head 840035c of matplotlib#18454. Benchmark replay, not for merge.