[Support]: shm based frame storage can go out of date and brick the relevant camera until shm buffers are cleaned up

[nagisa]

Checklist

• I have updated to the latest available Frigate version.
• I have cleared the cache of my browser.
• I have tried a different browser to see if it is related to my browser.
• I have tried reproducing the issue in incognito mode to rule out problems with any third party extensions or plugins I have installed.
• I have asked the AI at https://docs.frigate.video about my issue.

Describe the problem you are having

After an online system upgrade I started seeing error messages along the lines of ERROR : camera: Unable to read frames from ffmpeg process. The detect was not working, even though the stream was working and ffmpeg processes were producing the frames.

After some debugging and instrumentation I have found that the following exception was being raised

ValueError: memoryview assignment: lvalue and rvalue have different structures
[2026-08-21 19:27:11] frigate.video                  ERROR   : a: Unable to read frames from ffmpeg process.
Traceback (most recent call last):
  File "/nix/store/42d1qi92z9knq6l6l9xg2mgfw30zxgax-frigate-0.17.2/lib/python3.13/site-packages/frigate/video.py", line 141, in capture_frames
    frame_buffer[:] = ffmpeg_process.stdout.read(frame_size)
    ~~~~~~~~~~~~^^^

and that frigate was requesting frame_size = 1843200 bytes for a 1280x960 frame, but the actual shm was for a 1280x720 frame instead and sized 1382400. As a result no frame reads would ever succeed.

It seems to me that the incorrect frame size of 1280x720 gets detected (or defaulted to) when my homelab boots up from scratch after a power outage and then next time frigate starts, camera resolution is correctly detected as 1280x960. Unfortunately, even after detect.{width,height} are set, the shm size still sticks around as wrong.

Steps to reproduce

1. Prepopulate your /dev/shm with incorrectly sized camera_frame{1-32} shared buffers;
2. Start frigate;
3. Notice frigate attempting to reuse the precreated buffers.

Version

0.17.2 (although I believe it should also affect dev)

In which browser(s) are you experiencing the issue with?

No response

Frigate config file

audio:
  enabled: true
  listen:
  - bark
  - fire_alarm
  - scream
  - speech
  - yell
  - roar
  - growling
  - whimper_dog
  - crying
  - car_alarm
  - tire_squeal
auth:
  enabled: false
birdseye:
  enabled: false
cameras:
  a:
    ffmpeg:
      input_args: preset-rtsp-restream
      inputs:
      - path: rtsp://127.0.0.1:8554/a_record
        roles:
        - record
        - audio
      - path: rtsp://127.0.0.1:8554/a_detect
        roles:
        - detect
      output_args:
        record: preset-record-generic-audio-copy
    live:
      height: 960
      streams:
        Live: a_record
    motion:
      contour_area: 60
      threshold: 30
  c:
    audio:
      enabled: false
    detect:
      enabled: false
    ffmpeg:
      input_args: preset-rtsp-restream
      inputs:
      - path: rtsp://127.0.0.1:8554/c
        roles:
        - record
    live:
      height: 960
      streams:
        Live: c
    motion:
      enabled: false
  g:
    ffmpeg:
      input_args: preset-rtsp-restream
      inputs:
      - path: rtsp://127.0.0.1:8554/g_record
        roles:
        - record
        - audio
      - path: rtsp://127.0.0.1:8554/g_detect
        roles:
        - detect
      output_args:
        record: preset-record-generic-audio-copy
    live:
      height: 960
      streams:
        Live: g_record
    motion:
      contour_area: 60
      mask: 0.276,0.034,0.264,0.135,0.471,0.148,0.764,0.22,0.781,0.171,0.682,0.112,0.551,0.058,0.422,0.044
      threshold: 25
  d:
    ffmpeg:
      input_args: preset-rtsp-restream
      inputs:
      - path: rtsp://127.0.0.1:8554/d_record
        roles:
        - record
        - audio
      - path: rtsp://127.0.0.1:8554/d_detect
        roles:
        - detect
      output_args:
        record: preset-record-generic-audio-copy
    live:
      height: 960
      streams:
        Live: d_record
    motion:
      contour_area: 60
      threshold: 35
  mb:
    ffmpeg:
      input_args: preset-rtsp-restream
      inputs:
      - path: rtsp://127.0.0.1:8554/mb_record
        roles:
        - record
        - audio
      - path: rtsp://127.0.0.1:8554/mb_detect
        roles:
        - detect
      output_args:
        record: preset-record-generic-audio-copy
    live:
      height: 960
      streams:
        Live: mb_record
    motion:
      contour_area: 50
      mask: 0,0,0,0.488,0.862,0.105,0.88,0
      threshold: 40
  b:
    ffmpeg:
      input_args: preset-rtsp-restream
      inputs:
      - path: rtsp://127.0.0.1:8554/b_record
        roles:
        - record
        - audio
      - path: rtsp://127.0.0.1:8554/b_detect
        roles:
        - detect
      output_args:
        record: preset-record-generic-audio-copy
    live:
      height: 960
      streams:
        Live: b_record
    motion:
      contour_area: 50
      threshold: 40
  al:
    ffmpeg:
      input_args: preset-rtsp-restream
      inputs:
      - path: rtsp://127.0.0.1:8554/al_record
        roles:
        - record
        - audio
      - path: rtsp://127.0.0.1:8554/al_detect
        roles:
        - detect
      output_args:
        record: preset-record-generic-audio-copy
    live:
      height: 1280
      streams:
        Live: al_record
    motion:
      contour_area: 60
      frame_height: 320
      mask: 0,0,0.8,0.109,0.798,0.05,0.422,0
      threshold: 30
  t:
    ffmpeg:
      input_args: preset-rtsp-restream
      inputs:
      - path: rtsp://127.0.0.1:8554/t_record
        roles:
        - record
        - audio
      - path: rtsp://127.0.0.1:8554/t_detect
        roles:
        - detect
      output_args:
        record: preset-record-generic-audio-copy
    live:
      height: 960
      streams:
        Live: t_record
    motion:
      contour_area: 60
      threshold: 25
database:
  path: /var/lib/frigate/frigate.db
detect:
  enabled: true
  fps: 5
  max_disappeared: 300
  stationary:
    interval: 15
    threshold: 25
detectors:
  ov:
    device: GPU
    type: openvino
ffmpeg:
  hwaccel_args: preset-vaapi
  path: /nix/store/8s3h27dpxqqbvb6by2v5c3lfj7vs8v8y-ffmpeg-headless-8.1.1-bin
go2rtc:
  api:
    listen: 127.0.0.1:1984
    origin: '*'
  ffmpeg:
    bin: /nix/store/8s3h27dpxqqbvb6by2v5c3lfj7vs8v8y-ffmpeg-headless-8.1.1-bin/bin/ffmpeg
    global: -hide_banner -v error -vaapi_device /dev/dri/renderD128
  rtsp: {}
  streams:
    a_detect: rtsp://user:pass@a.local/live1s2.sdp
    a_record:
    - ffmpeg:rtsp://user:pass@a.local/live1s1.sdp#video=copy#audio=aac
    - ffmpeg:a_record#audio=opus
    c: ffmpeg:http://192.168.110.158:8080/?action=stream#video=h264#hardware
    g_detect: rtsp://user:pass@g.local/live1s2.sdp
    g_record:
    - ffmpeg:rtsp://user:pass@g.local/live1s1.sdp#video=copy#audio=aac
    - ffmpeg:g_record#audio=opus
    d_detect: rtsp://user:pass@d.local/live1s2.sdp
    d_record:
    - ffmpeg:rtsp://user:pass@d.local/live1s1.sdp#video=copy#audio=aac
    - ffmpeg:d_record#audio=opus
    mb_detect: rtsp://user:pass@mb.local/live1s2.sdp
    mb_record:
    - ffmpeg:rtsp://user:pass@mb.local/live1s1.sdp#video=copy#audio=aac
    - ffmpeg:mb_record#audio=opus
    b_detect: rtsp://user:pass@b.local/live1s2.sdp
    b_record:
    - ffmpeg:rtsp://user:pass@b.local/live1s1.sdp#video=copy#audio=aac
    - ffmpeg:b_record#audio=opus
    al_detect: rtsp://user:pass@al.local/live1s2.sdp
    al_record:
    - ffmpeg:rtsp://user:pass@al.local/live1s1.sdp#video=copy#audio=aac
    - ffmpeg:al_record#audio=opus
    t_detect: rtsp://user:pass@t.local/live1s2.sdp
    t_record:
    - ffmpeg:rtsp://user:pass@t.local/live1s1.sdp#video=copy#audio=aac
    - ffmpeg:t_record#audio=opus
  webrtc:
    candidates:
    - 192.168.100.1:8555
model:
  height: 320
  input_pixel_format: rgb
  input_tensor: nhwc
  labelmap_path: /var/lib/frigate/coco-80-animal.txt
  model_type: yolo-generic
  path: /var/lib/frigate/models/yolov9-c-320.onnx
  width: 320
motion:
  contour_area: 20
  frame_height: 240
  improve_contrast: false
  lightning_threshold: 0.3
  threshold: 30
mqtt:
  enabled: true
  host: mqtt.local
objects:
  filters:
    animal:
      min_score: 0.45
      threshold: 0.6
    car:
      min_score: 0.45
      threshold: 0.6
    person:
      min_score: 0.45
      threshold: 0.6
  track:
  - person
  - bicycle
  - car
  - motorcycle
  - animal
record:
  alerts:
    post_capture: 10
    pre_capture: 10
    retain:
      days: 15
      mode: active_objects
  detections:
    post_capture: 10
    pre_capture: 10
    retain:
      days: 10
      mode: motion
  enabled: true
  events:
    post_capture: 9
    pre_capture: 9
  retain:
    days: 5
    mode: all
review:
  alerts:
    labels:
    - animal
    - person
  detections:
    labels:
    - animal
    - person
    - bicycle
    - car
    - motorcycle
snapshots:
  enabled: false
telemetry:
  stats:
    amd_gpu_stats: false
    intel_gpu_stats: false
  version_check: false

docker-compose file or Docker CLI command

N/A

Relevant Frigate log output

[2026-08-21 19:27:11] frigate.video                  ERROR   : a: Unable to read frames from ffmpeg process.
(and so on for all cameras)

Relevant go2rtc log output

No output from go2rtc

Operating system

Other Linux

Install method

NixOS module

Network connection

Wired

Camera make and model

Vivotek 9380-H

Screenshots of the Frigate UI's System metrics pages

N/A

Any other information that may be helpful

Clanker identifies

frigate/frigate/util/image.py

Lines 1065 to 1076 in 036bae4

	def create(self, name: str, size) -> AnyStr:
	try:
	shm = UntrackedSharedMemory(
	name=name,
	create=True,
	size=size,
	)
	except FileExistsError:
	shm = UntrackedSharedMemory(name=name)
	self.shm_store[name] = shm
	return shm.buf

as the faulty code. Namely because the relevant shm file exists already, it just opens it up without checking the size or warning about it, instead of trying to fix it up in any way.

At minimum it should emit a warning on startup, but ideally IMO it should try to recreate the file so that frigate can be confident that the file is of correct size.

[inkeep-ai-assistant[bot]]

This behavior has been reported before with the same symptoms: after a detect resolution change, existing /dev/shm/<camera>_frame* segments remain at the old size, so reads misalign and the camera logs Unable to read frames from ffmpeg process indefinitely while the detect ffmpeg process stays healthy.
(1), (2)

The identified code path is the same one you point to — the frame slots are pre-created per camera:

for i in range(count):
    frame_size = config.frame_shape_yuv[0] * config.frame_shape_yuv[1]
    self.frame_manager.create(f"{config.name}_frame{i}", frame_size)

(2)

and create() falls back to opening the existing segment without validating its size:

try:
    shm = UntrackedSharedMemory(name=name, create=True, size=size)
except FileExistsError:
    shm = UntrackedSharedMemory(name=name)

(2)

which then fails at:

frame_buffer[:] = ffmpeg_process.stdout.read(frame_size)

(2)

Workaround: delete the camera's stale segments and restart so they are recreated at the correct size.

rm -f /dev/shm/<camera> /dev/shm/out-<camera> /dev/shm/<camera>_frame*

(1)

Or, for a systemd-style install:

systemctl kill --kill-whom=all frigate
rm -f /dev/shm/Office_frame*

systemctl reset-failed frigate
systemctl start frigate

(2)

Relevant context on your install method: /dev/shm is a path that should only live inside of the container, and a restart should clear everything in it — persistence of /dev/shm across restarts happens in non-container runtimes (e.g. running from systemd), which is where the stale ring survives.
(2)

The suggested fix matches your proposal: on camera start, compare the size of existing /dev/shm/<camera>_frame* segments against the expected detect_width * detect_height * 1.5 and unlink + recreate when the size differs, instead of reusing stale segments.
(1), (2)

For reference on sizing, /dev/shm is the internal cache for raw decoded frames in shared memory, and its minimum size is impacted by the shm-size calculation.
(3)

📚 Sources:

1. Changing a camera's detect resolution leaves stale /dev/shm/_frame* buffers that are never recreated, permanently breaking the camera with "Unable to read frames"
2. Frigate reuses stale per-camera SHM frame segments after detect resolution change
3. Installation

---

Have another question? Just tag @inkeep.

[hawkeye217]

The AI is correct here, /dev/shm is supposed to be internal to the container and go away when the process does, and Frigate already unlinks these segments on a clean shutdown. A normal docker install doesn't hit this. It's specific to your NixOS setup, so you'll want to scope /dev/shm to the unit or clear /dev/shm/*_frame* before start.

[nagisa]

Yeah, alright, I'll look for a way to address this on the systemd service end. I looked (asked the clanker) for the prior issues a few times, but it only found some after I started the discussion. Thank you.

[hawkeye217]

Right, maybe some kind of PrivateTmp style isolation or even just an rm -f in ExecStartPre.

Happy to help!

[nagisa]

In case anybody is looking for a solution for this in the future, I used TemporaryFileSystem = "/dev/shm:mode=1777,nosuid,nodev"; alongside PrivateTmp as implemented in NixOS/nixpkgs#555151.