14. Missing metadata/overlay on some frames with vvas_xmetaaffixer (tee back-pressure) - 14. Missing metadata/overlay on some frames with vvas_xmetaaffixer (tee back-pressure) - 6.3 English - XD056

In a tee + vvas_xmetaaffixer pipeline, some frames on the slave/overlay branch are output without metadata (for example, missing classification labels or bounding boxes) - typically the first few frames, until the affixer timeout elapses. The textual/master output might look correct while the overlay branch does not.

Root cause

The tee splits a common source into a master (inference) branch and one or more slave branches. If a slave-branch queue fills and back-pressures the tee, the tee blocks all of its branches - including the master branch that feeds vvas_xinfer. With no buffer arriving on the master pad, the affixer has no metadata to attach; its timeout watchdog then fires and force-pushes the slave buffers downstream without metadata.

This is most common with raw (uncompressed) video, where a single buffer can be large (for example a full-resolution RGB/RGBx or NV12 frame). The default queue byte limit (max-size-bytes, ~10 MB) can be smaller than one such buffer, so the queue fills after a single frame and immediately back-pressures the tee.

Confirm with debug logs

Enable the vvas_xmetaaffixer logs at level 7 (cumulative - captures both following messages):

GST_DEBUG=vvas_xmetaaffixer:7 gst-launch-1.0 ...

Two messages indicate the problem:

• Timeout occured - GST_INFO (level 4); the timeout watchdog fired because buffers did not arrive on all sink pads within the timeout duration.

• buffer is not availale at master pad - GST_LOG (level 7); the master (inference) branch has no buffer; that is, it is being starved.

Example log when the slave-branch queue back-pressures the tee:

0:00:02.003456789  1234  0xaaaabbbb INFO   vvas_xmetaaffixer gstvvas_xmetaaffixer.c:timeout_func:<ma> Timeout occured
0:00:02.003457012  1234  0xaaaabbbb LOG    vvas_xmetaaffixer gstvvas_xmetaaffixer.c:vvas_xmetaaffixer_process:<ma> buffer is not availale at master pad

Seeing Timeout occured together with repeated buffer is not availale at master pad messages - and output frames missing metadata until the timeout elapses - confirms the slave-branch queue is back-pressuring the tee.

Fix

Size the slave-branch queue so it does not back-pressure the tee. For raw video, remove the byte cap (and, if needed, the time cap):

... ! queue max-size-bytes=0 max-size-time=0 ! ma.sink_slave_0 ...

A large finite cap such as max-size-bytes=67108864 (64 MB) works equally well. These settings keep slave buffers queued long enough for the affixer to match them to the master metadata within the timeout window, so all output frames are correctly annotated.