Skip to main content
A plugin runs on a small SBC alongside the rest of the agent. The supervisor budgets resources tightly. This page covers what the budgets are, how to spot when you blow them, and how to profile.

​
Default budgets

When a plugin’s manifest does not declare resources, it gets:
ResourceDefault cap
RAM64 MB
CPU10% of one core
PIDs4
Disk100 MB
Inbound topic outbox256 messages per topic
Declare higher caps in the manifest: 
agent:
  resources:
    max_ram_mb: 256
    max_cpu_percent: 30
    max_pids: 8
    max_disk_mb: 500
Honest declarations matter. The operator sees them in the install dialog with a yellow warning if any of them are above the defaults. Lying to skip the warning ends with the operator disabling your plugin when it crashes the drone.

​
Spotting cgroup throttling

max_cpu_percent is enforced by the cgroup cpu.max controller. A plugin that breaches it does not crash; it gets throttled. The symptom is “everything got slower for no reason.” Read the throttle stats: 
cat /sys/fs/cgroup/ados-plugins.slice/ados-plugin@<id>.service/cpu.stat
The interesting field is nr_throttled. If it climbs every second, your plugin is over budget. Either optimize the code or declare a higher max_cpu_percent. ados plugin profile <id> (see below) surfaces this in a human-readable table.

​
Memory pressure

max_ram_mb is enforced by memory.max. A breach OOM-kills the process. The supervisor marks it crashed and the circuit breaker counts toward the three-strikes-in-five-minutes rule. Watch for the warning band: 
cat /sys/fs/cgroup/ados-plugins.slice/ados-plugin@<id>.service/memory.current
cat /sys/fs/cgroup/ados-plugins.slice/ados-plugin@<id>.service/memory.events
memory.events shows low, high, max, oom, oom_kill counters. Any non-zero high or max means the kernel started applying pressure (reclaiming page cache, swapping if allowed). To stay under the cap:
  • Process events on a bounded queue, not an unbounded list.
  • For ML models, prefer quantized weights (int8 over fp32).
  • Stream files; do not load whole logs into RAM.
  • Use array.array or numpy over a Python list of floats.

​
The ados plugin profile subcommand

ados plugin profile <id> runs for 60 seconds and prints a summary: 
plugin           com.example.battery
window           60s
cpu_user_pct     7.8 (cap: 10.0)
cpu_throttle_ms  120 (8 throttle events)
ram_avg_mb       38 (cap: 64)
ram_peak_mb      52
ram_events       0 high, 0 max, 0 oom
disk_used_mb     12 (cap: 100)
ipc_in_msgs      4400  (73/s)
ipc_out_msgs     280   (4.7/s)
ipc_inflight_p99 18 ms
back_pressure    0 events
restarts         0
Use the cap columns to see how close you are to your declared budgets. A plugin running consistently above 80% of any cap is one bad input away from a crash.

​
Profiling Python plugins

The SDK does not bundle a profiler; use cProfile or pyinstrument. The simplest path: 
import cProfile, pstats

def on_start_profiled(self, ctx):
    pr = cProfile.Profile()
    pr.enable()
    try:
        return self._real_on_start(ctx)
    finally:
        pr.disable()
        pstats.Stats(pr).sort_stats("cumulative").print_stats(30)
Or push an async handler through pyinstrument: 
uv pip install pyinstrument
python -m pyinstrument -m ados.plugin_runner com.example.battery
Output goes to journal under the unit name. Read it with journalctl -u ados-plugin@com.example.battery.service.

​
Profiling TypeScript plugins

Open the GCS in developer mode (?dev=1), pick the plugin’s iframe in the browser devtools (Sources tab), and use the Performance recorder. The iframe is a real browsing context with the standard devtools surface. Common GCS hot paths:
  • React re-renders triggered by every telemetry event. Coalesce with requestAnimationFrame and a local ref.
  • Heavy SVG redraws. Switch to canvas if the path count goes above ~200.
  • Synchronous JSON parse in the message handler. Move it to a Web Worker or off the main thread.

​
Event-loop discipline

The biggest cause of agent-plugin crashes is a blocked event loop. Symptoms: missed watchdog ping, supervisor SIGKILL, the crashed lifecycle event, and journal showing Watchdog timeout (limit 30s). Don’t:
  • time.sleep(...) in async code.
  • Heavy CPU work directly in on_start.
  • Synchronous I/O on a slow disk.
Do:
  • await asyncio.sleep(...).
  • await asyncio.to_thread(heavy_work).
  • aiofiles or chunked reads for slow I/O.

​
Disk discipline

Disk I/O is the second-most-common cause of throttling. A plugin writing 1 MB samples per second to its data dir on an SD card is already at the SBC’s sustained-write ceiling for the cheap end of the SD card market. To stay under the cap:
  • Batch writes. One fsync per second is plenty.
  • Compress logs (zstd is fast and saves a lot on JSONL).
  • Rotate. The SDK helper ctx.logging.rotating_file_handler does what it says.
  • Delete on uninstall. The host wipes your data dir, but in-life cleanup is your job.

​
Network discipline

Plugins with network.outbound are under no host-imposed bandwidth cap, but the drone’s link is shared with telemetry and video. A plugin pulling weights from the cloud during flight is a UX bug. Pull during ground time and cache.

​
Thermal behavior

On the Rock 5C Lite at 30C ambient, sustained 80% CPU on one core lifts the SoC about 18C. The agent throttles its own non-essential tasks at 75C. A plugin pushing the SoC into thermal throttle is observable as cgroup throttle events on every plugin, not just yours, because the kernel is reducing the budget pool. Watch the system telemetry topic: 
async with ctx.events.subscribe("telemetry.system") as stream:
    async for sample in stream:
        if sample.temperature_c > 70:
            self.reduce_workload()
The well-behaved plugin reduces its workload before the kernel forces the issue.

​
See also