Issue after Frigate 14 upgrade - CPU 100%

[juampick]

Hello! I've been using Frigate within Hassio for the last 2+ years. Before with Frigate 13 everything was working fine. I loved Frigate 14 but after upgrade, I have started to have issues and basically CPU goes to 100% and hangs frigate/hassio. The only solution is actually rebooting server.
My setup is an Intel Nuc with i5 (8th generation) + 16gb ram DDR4 + ssd + intel coral for frigate + 8 cameras (almost all are Dahua). The installation type is over HassOS as an Addon.
I use go2rtc, detect with substream pretty slow res (640px - 5fps) and then record with full res (1080p or 2k).

Everything runs fine, until many movements/motion is detected in any camera and then things starts to run pretty bad and hangs.

Frigate is the responsible for this, cpu gets stuck at 100%, Coral also has an average of less than 10 ms of inference speed (when everything is without movement), but when things are bad goes to 300-500 ms.

My config is the following: (I will keep it with 1 camera for simplicity only, but is the same with 8).

Things that I tried: changing from hwaccel_args: preset-vaapi to hwaccel_args: preset-intel-qsv-h264 (same results).

mqtt:
  enabled: true
  host: <host>
  port: 1883
  user: <user>
  password: <pass>

ffmpeg:
 hwaccel_args: preset-intel-qsv-h264

detectors:
  coral:
    type: edgetpu
    device: usb

##########################################################################
cameras:

cameras:
  front:
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/front
          input_args: preset-rtsp-restream
          roles:
            - record
        - path: rtsp://127.0.0.1:8554/front-sub
          input_args: preset-rtsp-restream
          roles:
            - detect
    objects:
      track:
        - person
        - car
        - cat 

version: 0.14
##########################################################################
go2rtc:
  log:
    exec: trace
  streams:
    front:
      - rtsp://user:pass@ip:554/cam/realmonitor?channel=1&subtype=0
    front-sub:
      - rtsp://user:pass@ip:554/cam/realmonitor?channel=1&subtype=1
# ##########################################################################

snapshots:
  # Optional: Enable writing jpg snapshot to /media/frigate/clips (default: shown below)
  enabled: true
  timestamp: true

record:
  enabled: true
  retain:
    days: 7
    mode: motion
  events:
    retain:
      default: 30
      mode: motion

objects:
  track:
    - car
    - person
##########################################################################
version: 0.14

Can anyone help me? I have been struggling for the last 2 weeks or more trying to solve this issue.

Thanks ;)

[NickM-27]

the best thing to see would be the system and camera metrics

[juampick]

Exactly, same usb port, (3.0) but with usb 3.0 cable is not working, and it's working with usb 2.0 cable.
This was working smoothly with frigate 13.

frigate upgrade to 14 for me broke everything.

[NickM-27]

Nothing regarding the coral was changed in 0.14 and we've not had any other reports like this

[juampick]

Yes. It’s pretty weird. I am using HassOS. Do you know if changing to a Proxmox lxc and having containers per hassio, frigate, etc would be better?

[juampick]

I am not sure what other thing to try

[NickM-27]

More virtualization usually is worse but hard to say for sure

[mbronstein1]

I've noticed a similar issue on my end. I had 7 cameras on version 13 with no detect stream issues and relatively low CPU. Migrating to version 15 and I consistently am not able to access the detect streams for any of the cameras and cpu usage is pretty high (generally 85-100%). I am using the same number of cameras.

[hawkeye217]

Check out the breaking changes in the 0.15 beta 1 release notes. This is probably related to the upgrade to ffmpeg 7. If you change that and are still experiencing issues, go ahead and open your own support discussion.

[juampick]

UPDATE:

I was having pretty long issues with this after Frigate 0.14 upgrade. My setup was a HassOS installation baremetal on Intel Nuc.
I tried all the USB ports of the NUC. Several USB 3.0 cables without luck.

Last option was changing the installation type. I changed it to a Proxmox VE. Installed in one container, a VM with HassOS, and in a LXC, frigate using the default installation with the helper script.
I then did the USB passthrough, and after facing some issues with coral drivers I found the solution here. And that worked fine.
This is the amount of resources I assigned to that LXC:

This has been running for 55hs without issues. And look at the values, pretty impressive. Ah, I switched to the original USB 3.0 cable that came with the Google Coral. (That with previous setup was giving me inference speeds of up to 400-500ms).

So, believe it or not, seems that switching to Proxmox solved the issue. Weird, at least.
Another pro tip, is that using Proxmox and having HA in a separate container, will make not to break HA if something bad happens into Frigate.

Thanks @NickM-27 for your support.