Memory issue

[Chalkybot]

I'm using this library to listen to an online radio. My problem is that hte memory usage of the script goes up by around 1MB every minute or two. This leads to the process eating up potentially gigabytes of memory if I listen to the radio for long periods of time, as i do. I have tested and this does not occur without starting the "play" function.

[dfaker]

A quick test of this with these flags set:

import mpv

player = mpv.MPV()
player.cache=False
player.demuxer_max_bytes=100
player.demuxer_readahead_secs=0

player.play('https://media-ice.musicradio.com/CapitalMP3')
t = time.time()

Logging memory with

process = psutil.Process(os.getpid())
process.memory_info().rss

Playback time (s)	Total Memory (b)	Memory Increase since play (b)
5	37052416	4096
60	37163008	118784
120	37163008	118784
180	37437440	393216
240	37437440	393216
300	37978112	933888
540	37978112	933888
600	38014976	970752
810	38760448	1716224
860	38080512	1036288
872	33800192	-3244032
1043	34840576	-2203648
1200	33882112	-3162112
1320	33882112	-3162112
1440	33882112	-3162112
1500	33882112	-3162112
1560	33882112	-3162112
1620	33882112	-3162112
1680	33882112	-3162112
1740	33882112	-3162112

Tracemalloc shows no matching increase or decrease in python variable memory.

Process monitor seems to confirm a similar freeing.

So this would seem like normal caching and release on the mpv side, interestingly I'm seeing notable spikes when the tracks change or it switches to adverts so perhaps some kind of metadata in the stream is being cached?

Have we confirmed the predicted gigabytes of memory usage in practice or is it extrapolated?

If so is it reproducible with those cache flags set on the mpv instance?

[jaseg]

I'm reasonably sure that this is an issue inside libmpv, not python-mpv. First of all, there are not many places where python-mpv could leak memory while doing nothing. The only place where data gets passed between libmpv and python-mpv in the background is the event loop, but when running @dfaker's test code, the event loop does not receive any more events after the stream has started. Also, in its internal bookkeeping mpv uses a fixed-size buffer for the event loop and automatically frees the events (we don't have to do that from python).

The second indicator that this is a libmpv issue is that I observe the same issue when running command-line mpv on the stream URL from @dfaker's reply. From a cursory search it seems like this could be https://github.com/mpv-player/mpv/issues/5889 . On my system, mpv --demuxer-max-bytes=100k --demuxer-max-back-bytes=100k already seems to stop growing after a few seconds. From the mpv manpage's description of these options, beyond these two there are still other caches inside mpv's processing pipeline that might grow over time.

[jaseg]

I'm closing this as it seems that this is a libmpv issue. Please feel free to re-open or submit a new issue if it turns out that it is actually on the python-mpv side.