search

nvidia-holoscan / holoscan-sdk

The AI sensor processing SDK for low latency streaming workflows
Apache License 2.0
117 stars 31 forks source link

VideoStreamRecorderOp fails create *.gxf_index file on the flight #22

Closed jpmazellier closed 8 months ago

jpmazellier commented 8 months ago

I posted a question on the NVIDIA developer forum about this issue: [(https://forums.developer.nvidia.com/t/videostreamrecorderop-issue-with-gxf-index-empty/285639)] Basically, the .gxf_entities file is well populated by the operator VideoStreamRecorderOp on the flight, but not the .gxf_index file (empty) until the application closes "properly". The only way i found was to have a "proper" application closing (using a VideoStreamReplayerOp with a fixed number of frames as source video). So in case of crash or forced interruption, the video stream can not be reconstructed. Would it be possible to have the *.gxf_index file populated regularly so to have a chance to recover the video even in case of application crash ? Thanks !

tbirdso commented 8 months ago

Hi @jpmazellier , thank you for reporting this issue and documenting your experience in the Holoscan forum! We will look into what it would take to update VideoStreamReplayerOp behavior as you've described.

tbirdso commented 8 months ago

Hi @jpmazellier , after investigation we've determined that this feature is currently available in VideoStreamRecorderOp in Holoscan v1.0.3. Please set the flush_on_tick option to True when initializing your VideoStreamRecorderOp in order to flush both the binary and index streams on each tick. Note that you may observe a performance impact with this feature enabled.

I will close this issue as already implemented, but please re-open it if you continue to observe issues with flush_on_tick enabled. If you have further questions on operator parameters or VideoStreamRecorderOp please add them to your forum thread and we will follow up there.

Thank you for using Holoscan!