fix(mpl): contain comm callback errors in mpl_interactive

[kirangadhave]

Summary

A raising callback inside mpl_interactive._handle_comm_msg (e.g. matplotlib's AttributeError: 'Axes' object has no attribute '_pan_start' from end_pan without a prior start_pan) was propagating up through MarimoComm.handle_msg → WIDGET_COMM_MANAGER.receive_comm_message → the kernel's request handler, terminating asyncio.run and killing the kernel subprocess. The uvicorn parent kept serving so the UI just reported "stopped responding".

This wraps the handle_json and _handle_download calls in try/except Exception and logs via LOGGER.exception so a buggy callback is contained while the traceback still surfaces for diagnosis. BaseException is intentionally not caught.

Fixes the kernel-crash half of the matplotlib pan/zoom issue (see linked Linear issue). The underlying matplotlib _pan_start bug is tracked separately.

Test plan

• Manual repro via the linked Linear repro (pan → re-run cell → zoom → pan): exception logged, kernel stays alive, figure remains responsive.
• Direct programmatic injection: _handle_comm_msg with a raising handle_json returns cleanly and a follow-up message is processed.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
marimo-docs	Ready	Preview, Comment	May 12, 2026 10:40pm

[cubic-dev-ai]

No issues found across 1 file

Architecture diagram

sequenceDiagram
    participant UI as Browser UI
    participant UV as Uvicorn Server
    participant KRN as Kernel Subprocess
    participant CM as WIDGET_COMM_MANAGER
    participant MC as MarimoComm
    participant MPI as mpl_interactive
    participant FM as Figure Manager
    participant LOG as Logger

    Note over UI,LOG: Matplotlib Pan/Zoom Event Flow

    UI->>UV: Comm message (pan/zoom event)
    UV->>KRN: receive_comm_message()
    KRN->>CM: handle message
    CM->>MC: handle_msg()
    MC->>MPI: _handle_comm_msg()

    alt Event type is "download"
        MPI->>MPI: _handle_download(fmt)
        alt Download succeeds
            MPI->>FM: savefig()
            MPI->>UI: Send blob via comm
        else Exception in download
            MPI->>LOG: LOGGER.exception()
            MPI->>MPI: Return cleanly
        end
    else Event type is JSON (mouse/toolbar)
        MPI->>FM: handle_json(event)
        alt Callback succeeds
            FM-->>MPI: void
            MPI->>UI: Updated figure state
        else Exception propagates
            FM-->>MPI: raise Exception
            Note over MPI,LOG: CHANGED: Exception caught here
            MPI->>LOG: LOGGER.exception()
            MPI->>MPI: Continue processing
        end
    end

    alt Exception not caught (before fix)
        MPI->>KRN: Unhandled exception propagates
        KRN->>KRN: asyncio.run terminates
        KRN-->>UV: Subprocess dies
        UV->>UI: "stopped responding"
    else Exception caught (after fix)
        KRN->>KRN: Remains alive
        KRN-->>UI: Normal response continues
    end

    Note over UI,LOG: Error handling boundary: all exceptions from FM callbacks<br/>are contained within mpl_interactive layer

Loading

[Light2Dark]

should we add tests?

[Copilot]

Pull request overview

This PR prevents exceptions raised inside matplotlib interactive comm callbacks from propagating through the comm manager and crashing the kernel subprocess, while still emitting tracebacks for debugging.

Changes:

• Wrap self._figure_manager.handle_json(event) in try/except Exception and log failures with LOGGER.exception.
• Wrap figure download rendering/sending logic in try/except Exception and log failures with LOGGER.exception.

[github-actions]

🚀 Development release published. You may be able to view the changes at https://marimo.app?v=0.23.7-dev12