Describe the bug
When a file URI is present in uri-list at startup (our "idle fill" keep-alive clip),
the pipeline does process frames end-to-end, but worker threads grow without bound until the
process is OOM-killed.
- Threads named
pool-multistream-pipeline grow continuously, mostly parked in
futex_do_wait (ls /proc/1/task | wc -l, top -H).
GST_DEBUG=4 shows nvmultiurisrcbin/multi-uri_creator posting ~172,000 LATENCY
messages per second, each triggering latency recalculation + partial state re-sync, in
a self-sustaining loop; pending pipeline state never settles.- A file with an audio track (
sample_1080p_h264.mp4) grows at ~85 threads/s; a
video-only clip grows at ~7 threads/s — audio aggravates but is not the cause.
- A live RTSP source in
uri-list at startup does not exhibit this — so the trigger
is specifically a file source at startup.
Steps/Code to reproduce bug
gst-launch-1.0 nvmultiurisrcbin port=9000 ip-address=localhost \
batched-push-timeout=33333 max-batch-size=4 live-source=1 drop-pipeline-eos=1 \
width=1920 height=1080 \
uri-list=file:///opt/nvidia/deepstream/deepstream/samples/streams/sample_1080p_h264.mp4 \
! fakesink sync=0
# observe thread count climbing and GST_DEBUG=4 LATENCY message storm
Environment details (please complete the following information):
• Hardware Platform (Jetson / GPU): GPU (x86_64, dGPU)
• DeepStream Version: 9.0
• TensorRT Version: 10.14.x (DS 9.0 default)
• NVIDIA GPU Driver Version: 595.71.05
• CUDA Version: 13.x (DS 9.0 default)
• Issue Type: bug
• Containerized: yes (DeepStream 9.0 base image), Kubernetes
Additional context
Is the latency-recalculation storm with a file source in uri-list a known issue, and is
there a property (e.g. disabling latency renegotiation, or a live-source/sync setting)
that prevents the unbounded thread growth?
By submitting this issue, you agree to follow our code of conduct and our contribution guidelines.
Describe the bug
When a file URI is present in
uri-listat startup (our "idle fill" keep-alive clip),the pipeline does process frames end-to-end, but worker threads grow without bound until the
process is OOM-killed.
pool-multistream-pipelinegrow continuously, mostly parked infutex_do_wait(ls /proc/1/task | wc -l,top -H).GST_DEBUG=4showsnvmultiurisrcbin/multi-uri_creatorposting ~172,000LATENCYmessages per second, each triggering latency recalculation + partial state re-sync, in
a self-sustaining loop; pending pipeline state never settles.
sample_1080p_h264.mp4) grows at ~85 threads/s; avideo-only clip grows at ~7 threads/s — audio aggravates but is not the cause.
uri-listat startup does not exhibit this — so the triggeris specifically a file source at startup.
Steps/Code to reproduce bug
Environment details (please complete the following information):
• Hardware Platform (Jetson / GPU): GPU (x86_64, dGPU)
• DeepStream Version: 9.0
• TensorRT Version: 10.14.x (DS 9.0 default)
• NVIDIA GPU Driver Version:
595.71.05• CUDA Version: 13.x (DS 9.0 default)
• Issue Type: bug
• Containerized: yes (DeepStream 9.0 base image), Kubernetes
Additional context
Is the latency-recalculation storm with a file source in
uri-lista known issue, and isthere a property (e.g. disabling latency renegotiation, or a
live-source/syncsetting)that prevents the unbounded thread growth?
By submitting this issue, you agree to follow our code of conduct and our contribution guidelines.