Running time.sleep_ms(10) in tight loops

[davefes]

Copilot came up with this explanation when running a tight loop:

In MicroPython:

time.sleep_ms(ms) does yield to the MicroPython scheduler
It also allows the underlying FreeRTOS tasks to run
It prevents the VM from locking up

It keeps Wi‑Fi and internal housekeeping alive

So the effect is the same:
A short sleep keeps the system responsive and prevents watchdog resets.
🧩 Why this matters even more in MicroPython

MicroPython runs:

a Python VM
a garbage collector
a cooperative scheduler
plus all the ESP32 background tasks

If you write a tight loop like:
python

while True:
do_something()

…you can easily:

starve the garbage collector
starve Wi‑Fi
starve the idle task
trigger the watchdog
freeze the VM

So MicroPython developers often use:
python

time.sleep_ms(10)

or even:
python

await asyncio.sleep_ms(0)

to yield control.

Well, I have been working with the ESP32 for 5+ years and had never heard this before. It seems to be a major issue when working with an RTOS. Is this mentioned somewhere in the docs?

[scruss]

I think it's Copilot making stuff up again. I'd never heard or experienced this before.

The recommended "nice" function to use in a tight loop is machine.idle()

[davefes]

Looking at the link I see "to reduce power consumption". No mention of "the wheels falling off" if you don't use it!

[scruss]

Angus's answer is much more useful, though. And I'm very much feeling:

(The problem with our new exciting AI era is that it's taken ten minutes and my tired brain to write this answer, whereas Copilot can produce a short essay doubling down with even more nonsense in just a few seconds.)

[projectgus]

I'm afraid the main issue here is that AI is confidently telling you things which are almost entirely incorrect. (It's quite good at that, unfortunately.)

• "starve the garbage collector" -> No. MicroPython doesn't do any garbage collection in the background. Garbage collection happens either when you call gc.collect() or when some Python code tries to allocate memory.
• "starve Wi-Fi" -> No. Wi-Fi runs in other RTOS tasks which have higher priority than the MicroPython task. On dual core ESP32s there is even a whole second core to run these tasks!
• "starve the idle task" -> This is technically true, in the sense that if the system isn't idle then the idle task doesn't run. I can't think of a case where this would matter (maybe if you create and delete a lot of threads, because the memory from a deleted thread is reclaimed by the idle task, but that's stretching it.).
• "trigger the watchdog" -> No. There are watchdogs on ESP32 but none of them are triggered just by running Python code without sleeping.
• "freeze the VM" -> Having trouble conceiving how that would work.

(The problem with our new exciting AI era is that it's taken ten minutes and my tired brain to write this answer, whereas Copilot can produce a short essay doubling down with even more nonsense in just a few seconds.)

It is generally good to do a time.sleep_ms() or call machine.idle() if you know that your code is waiting for something in a loop. That part is true! The benefits are a lot less dramatic though - a small reduction in power usage, and on esp32 if a "soft interrupt" triggers then calling one of these functions means it'll be able to run immediately instead of a few hundred microseconds later if Python code was executing..

Is this mentioned somewhere in the docs?

Most of the parts which are real are, I think.

[GitHubsSilverBullet]

It's Dunni^H^H^H^H^HStunning, says Kruger!

[davefes]

What I find "entertaining" is when it makes a mistake it will say good-catch and then try harder to explain the nuances of its choices. Has anyone had a "sorry and will try harder next time"?

[Josverl]

I'm afraid the main issue here is that AI is confidently telling you things which are almost entirely incorrect. (It's quite good at that, unfortunately.)

And that as a responsible human you need to challenge everything that you can't explain based on your own understanding or research.

I refer to that proces as learning 🧐

[Josverl]

use NI

As in: "The knights that say NI" ?

[robert-hh]

Has anyone had a "sorry and will try harder next time"?

I had that sometimes. Until now I never got a useful response to difficult problems. AI can provide good answers, if the definitive right answers exist in the Internet, and the benefit of the AI is to find them in any of the supported natural languages and to proved the answer in your language in good style. LLM. Any of my trial attempt to create code besides "Hello World" level produced code that could not even get compiled.