RSDK-13042 — video-store using GetImages continues even if input is invalid repeatedly

[hexbabe]

RSDK-13042

Bug Summary

Steps to reproduce :

Use depth camera e.g. Realsense and hook it up as the underlying camera to a video-store.

Configure the camera to return only a color image at first and store some video, then change it to return only a depth map in its GetImages response e.g. RealSense.

Expected behavior :

Returns depth mime type invalid error, and stops storing video.

Actual behavior :

Returns depth mime type invalid error, but continues storing stale color frame as video.

Manual Tests

Config:

{
  "components": [
    {
      "name": "vs",
      "api": "rdk:component:camera",
      "model": "viam:video:storage",
      "attributes": {
        "camera": "oak-d",
        "sync": "data_manager-1",
        "storage": {
          "size_gb": 1
        }
      },
      "depends_on": [
        "data_manager-1"
      ]
    },
    {
      "name": "oak-d",
      "api": "rdk:component:camera",
      "model": "viam:luxonis:oak-d",
      "attributes": {
        "sensors": [
          "color"
        ]
      }
    }
  ],
  "services": [
    {
      "name": "data_manager-1",
      "api": "rdk:service:data_manager",
      "model": "rdk:builtin:builtin",
      "attributes": {
        "additional_sync_paths": [],
        "sync_interval_mins": 0.1,
        "capture_dir": "",
        "tags": []
      }
    }
  ],
  "modules": [
    {
      "type": "registry",
      "name": "viam_oak",
      "module_id": "viam:oak",
      "version": "latest"
    },
    {
      "type": "local",
      "name": "local-vs",
      "executable_path": "/Users/seanyu/Projects/sean-video-store/bin/darwin-arm64/video-store"
    }
  ]
}

Verified video saves well. Change sensors from ["color"] to ["depth]. Observed:

3/12/2026, 1:45:28 PM warn local-vs.rdk:component:camera/vs     videostore/videostore.go:583 expected image mime type image/jpeg got image/vnd.viam.dep:   log_ts 2026-03-12T17:45:28.138Z

and no more video saved.

After changing back to ["color"] there is one corrupted video, but the video resource recovers and keeps writing segments.

PR

2 reviewers

This was not vibe coded sorry for the slop earlier

TODO: make an RC, bump viamrtsp, and consume this change there

[seanavery]

Manual tests looking good, but wondering about the case in which the camera GetImages endpoint recovers back to color within a segment boundary. Does the encoder and resulting video file recover?

[seanavery]

Should we add a test case here for valid jpeg mime_type but invalid image bytes? to make sure we clear in that case

[hexbabe]

yes I will add.

[hexbabe]

looked into it: there's no explicit path in our code that checks and propagates a sentinel error we own on bad bytes. Want me to add that? Or would you rather assert on err string from underlying lib?

[SebastianMunozP]

[nit] We may get a small time win here if we only store a nil byte if its content is not nil already (avoiding the atomic operations). But totally fine the way this is.

[hexbabe]

done

[hexbabe]

Manual tests looking good, but wondering about the case in which the camera GetImages endpoint recovers back to color within a segment boundary. Does the encoder and resulting video file recover?

yes I actually tested this but forgot to include in the writeup. Thanks for asking!

edit: Description updated.