mediapane-relay

A mediapane plugin for clusters whose batch jobs can't reach your pane.

Compute nodes usually can't see the login node's /tmp, so a mediapane ssh socket doesn't help a job running under Slurm. Instead, give the job a spool folder on a filesystem both can see:

# in the job script (or your shell rc on the cluster)
export MEDIAPANE_SPOOL=~/.cache/mediapane/spool
python train.py          # mediapane.show(...) / mp ... as usual

When no pane is reachable, every push is written to that folder instead (one .mp file per push, in the wire format). On a machine that can reach the pane, forward them:

mediapane ssh login-node      # from your laptop: the pane's socket is forwarded
mediapane relay               # on the login node, inside that session (say, in tmux)

mediapane relay checks the folder every --interval seconds (0.5 by default), sends complete pushes oldest first, and deletes each once the pane has accepted it. Pushes the pane rejects are moved to failed/ in the spool. If no pane is reachable it says so once and keeps them until one is. Partial files left by writers that died are removed after an hour. --once forwards what's there and exits (status 1 if a pane couldn't be reached), for cron or a job epilogue.

Options: --spool DIR (default $MEDIAPANE_SPOOL), --interval SECONDS, --once.

Install

uv tool install mediapane --with mediapane-relay

Only the machine running the relay needs the plugin. Jobs only need mediapane itself.

Development: uv sync && uv run pytest in this directory.