liveplot.warm(): keep a render process ready so plots appear immediately

[davidquarel]

Example

import liveplot
from liveplot import LivePlot

liveplot.warm()        # once, in a setup cell: a render process is kept ready in the background

plot = LivePlot(loader, "loss | acc")     # ... so this plot's first frame appears right away
for batch in plot:
    plot.log(loss=train_step(batch))

What

A plot's first frame used to wait ~0.75 s for its render process, and nearly all of that is a fresh Python importing matplotlib (process start 0.04 s, import and figure 0.55 s, first draw 0.05 s). The render worker now does that warm-up first and then waits for its configuration, so a process can be started before any plot exists.

liveplot.warm() starts one. Each LivePlot adopts the waiting process and a replacement starts in the background, so every later plot is fast too. It's opt-in: without warm() nothing changes and no process is kept. The spare ignores interrupts and exits when the kernel dies, like the renderer always has.

The tour's setup cell (examples/demo.py) now calls it, and the guide and API pages document it.

Measured

In a real kernel via nbclient, three plots in a row with a 2 s gap between cells (someone reading the notebook):

	plot 1	plot 2	plot 3
without warm()	0.76 s	0.72 s	0.78 s
with warm()	0.21 s	0.24 s	0.21 s

If cells run back to back with no gap, a plot can adopt a replacement that is still warming up and gets part of the benefit (measured 0.59 s).

No demo GIF: a recording starts at the first frame, so the wait this removes can't be seen in one; the table is the evidence.

Verified

tests/test_warm.py: warm() is exported and idempotent; a plot adopts the spare and a replacement starts; the warm first frame is under 0.4 s; a dead spare is replaced rather than used; an unused spare exits when its parent dies; and without warm() plots leave no process behind. Full suite: 39 passed. The tour runs end to end as a script; the site builds strict-clean.

Stacked on #11.

🤖 Generated with Claude Code

https://claude.ai/code/session_01G9zs684RFxHvW6cNWqTPSF