Potential memory leak in _removed_tracks handling

Hobbes1987 · January 23, 2026, 9:17am

Please see:

Potential memory leak in _removed_tracks handling

opened 09:15AM - 23 Jan 26 UTC

Hobbes1987

In the current implementation, _removed_tracks grows unbounded over long-running… streams: `self._removed_tracks.extend(removed_stracks)` https://github.com/DeGirum/degirum_tools/blob/2fa19f009d5202ec19755f176e1d5539feab4bf8/degirum_tools/analyzers/object_tracker.py#L730 Each entry is a full STrack object, which includes Kalman filter state (mean, covariance) and other numpy arrays. In 24/7 deployments this leads to continuous memory growth (several KB per removed track), eventually making the tracker unusable on edge devices. From code inspection, _removed_tracks is only used in the same update cycle for: `self._lost_tracks = _ByteTrack._sub_tracks(self._lost_tracks, self._removed_tracks)` As far as i can see there is no functional dependency on historical removed tracks. Track IDs are monotonically increasing and never reused, and once a track is marked Removed it should not reappear in _lost_tracks in later frames. But i'm not sure of that. # Option A A safe and minimal fix is to assign instead of extend: `self._removed_tracks = removed_stracks` This keeps _removed_tracks limited to the current frame, avoids retaining heavy STrack objects indefinitely, and preserves identical behavior for _lost_tracks filtering in the same update step. ## Why this appears safe - track_id values are not reused - _removed_tracks is not referenced outside the current update cycle ## Consequence - No historical list of removed tracks is kept (which currently does not appear to be used anywhere) - Memory usage remains stable over long runt As implemented now, `_removed_tracks.extend` causes unbounded memory growth and is not suitable for long-running or embedded/edge use cases. # Option B Another option would be to use `deque` instead of a `List[STrack]` ``` from collections import deque self._removed_tracks = deque(maxlen=2000) # choose N ``` > Both options seem not to affect my ai analysis in any negative way

I’ll hereby post this on this community forum as well. I’m not sure what medium is preferred for bug reports.

shashi · January 23, 2026, 3:32pm

Hi @Hobbes1987

Thanks for bringing this to our attention. We will investigate and keep you posted.

vladk · January 27, 2026, 10:58pm

Hi @Hobbes1987 ,

We released degirum-tools ver. 1.0.0 with this fix.

To install:
pip install -i https://pkg.degirum.com degirum-tools

Note, that now degirum-tools is distributed from DeGirum packaging service (https://pkg.degirum.com).

Hobbes1987 · January 29, 2026, 1:19pm

Hi! Thanks for fixing this for us!

When i update to 1.0.0 my composition is not starting, when downgrading to degirum 0.20.0 and degirum-tools 0.26.1 my composition works.

i’m currently on a “jetson orin nano”. for now to proceed in my tests i’ll stick to this version combination.

shashi · January 29, 2026, 8:32pm

Hi @Hobbes1987

Can you please provide more details on the problem? Are you seeing any error messages? Also, did you upgrade degirum or degirum-tools or both?

Hobbes1987 · January 30, 2026, 1:48pm

I added these lines in my python code:
logging.getLogger("degirum").setLevel(logging.DEBUG)
logging.getLogger("degirum_tools").setLevel(logging.DEBUG)

Now i see:

DEBUG degirum.model: Traceback (most recent call last):
.../venv/lib/python3.10/site-packages/degirum/model.py", line 1230, in _predict_impl
    self._runtime.predict(preprocessed_data, "")
RuntimeError: [ERROR]Functionality is not supported
License does not allow usage of runtime agent 'TENSORRT'
dg_task_runner.cpp: 103 [DG::CoreTaskRunner::processorCreate]

I can use degirum-tools 1.0.0, with degirum 0.20.0 that works

vladk · February 3, 2026, 5:50pm

@Hobbes1987 ,

Please be advised that starting from ver.1.0.0, PySDK requires license for premium runtimes.
See more info here:

Release Notes | PySDK | DeGirum Docs

Basically, all you need is to install a workspace token: degirum token install --token <TOKEN>.
You need to create workspace on AI hub and then generate a token there.
This will give you one free installation of PySDK premium runtimes (node locked)

Instructions:
Workspaces | AI Hub | DeGirum Docs
Workspace Tokens | AI Hub | DeGirum Docs

Topic		Replies	Views
Two processes created by one script General pysdk , degirum-tools	6	85	October 28, 2025
EventDetector(), EventNotifier() and gotify General hailo , raspberry-pi , compiler , degirum-tools	59	361	January 15, 2026
Issue whiile running General	14	66	February 10, 2026
Getting TRAILS from ObjectTracker - need help! General	3	42	December 3, 2025
Share a quick testimonial Feedback pysdk , ai-hub , compiler , degirum-tools	1	35	November 24, 2025

Potential memory leak in _removed_tracks handling

Related topics