Simulation actors register process-global repeaters/handlers with no per-instance disposal (memory leak in per-request embeddings) #436

pip install rosys, then python leak.py. The page builds a RobotSimulation + Odometer (a 100 Hz and a 1 Hz repeater); each page-view adds both and never releases them:

import threading
import urllib.request

from nicegui import app, ui
import rosys                 # noqa: F401
import rosys.rosys as core

@ui.page('/')
def demo() -> None:          # a normal live "steer a simulated robot" demo
    from rosys.driving import Odometer, robot_object
    from rosys.geometry import Prism
    from rosys.hardware import RobotSimulation, WheelsSimulation
    wheels = WheelsSimulation()
    RobotSimulation([wheels])                                   # -> 100 Hz Repeater(self.step)
    with ui.scene():
        robot_object(Prism.default_robot_shape(), Odometer(wheels))  # -> 1 Hz Repeater(prune_history)

def drive() -> None:
    for i in range(1, 601):
        urllib.request.urlopen('http://localhost:8082/').read()   # one page-view
        if i % 100 == 0:
            live = len([t for t in core.Repeater.tasks if not t.done()])
            print(f'{i:3d} page-views -> {live} live Repeaters')
    import os
    os._exit(0)

app.on_startup(lambda: threading.Thread(target=drive, daemon=True).start())
ui.run(port=8082, reload=False, show=False)

Output — two repeaters added per page-view, never reclaimed:

100 page-views ->  202 live Repeaters
200 page-views ->  402 live Repeaters
400 page-views ->  802 live Repeaters
600 page-views -> 1202 live Repeaters

These loops are not reclaimed when the client goes away: I separately confirmed that after the NiceGUI clients disconnect and are pruned (Client.instances back to baseline), the live Repeater count stays elevated — client deletion tears down UI elements, not these global loops.

The bare simulation above leaks ~0.1 MB/visit, so RSS climbs slowly (and the event loop saturates at ~2200 leaked repeaters before it gets far). But a real driving/mapping robot carries state — a local occupancy grid, camera frames, etc. When each leaked robot holds a realistic ~5.5 MB local map, the same unbounded leak crosses 1 GB by ~160 page-views on my machine:

tracemalloc (measured separately from RSS — enabling it inflates RSS) attributes the ongoing allocation to the still-running leaked loops: rosys/driving/odometer.py:63 (self.history.append(...)) and rosys/geometry/pose.py.

rosys.on_repeat(handler, interval) returns a Repeater, but the actors that call it in __init__ discard the handle:

• rosys/hardware/robot.py — RobotSimulation.__init__ → on_repeat(self.step, 0.01)
• rosys/driving/steerer.py — on_repeat(..., 0.05)
• rosys/driving/odometer.py — on_repeat(self.prune_history, 1.0)

Repeater.start() adds the task to the class-global Repeater.tasks (rosys/rosys.py); it is removed only by Repeater.stop(), which callers can't reach because the handle was thrown away (Repeater.stop_all() cancels but doesn't remove). Deleting the NiceGUI client tears down its UI elements but not these global loops. Separately, Automator.__init__ appends a global rosys.on_shutdown(...) per instance, and Camera.__init__ (base of SimulatedCamera) registers global /images/<id>/… routes (same-id construction replaces the routes rather than accumulating them).