Fix LuxonisTracker robustness and add test coverage

[kozlov721]

Purpose

luxonis_ml/tracker had no tests. The module was also excluded from pyright and
from coverage, so neither tool looked at it. Real defects hid behind those
exclusions. This PR removes both exclusions, adds a test suite, and fixes the
defects they hid.

Specification

Breaking changes

Change	What a caller must do
LuxonisRequestHeaderProvider and luxonis_ml.tracker.mlflow_plugins are removed	The tracker no longer sends Cloudflare Access headers. MLFLOW_CLOUDFLARE_ID and MLFLOW_CLOUDFLARE_SECRET stay in environ, but no code reads them. Register your own MLflow request-header provider.
The public methods log_to_mlflow and store_log_locally are removed	Use the public log_* and upload_* methods. The private _mlflow_call replaces both.
local_logs is now list[BufferedCall]	It was a dict of the groups metric, metrics, params, images, artifacts, and matrices. Code that reads the buffer must change.
upload_artifact_to_mlflow calls mlflow.log_artifact	MLflow stores the artifact under its own file name. name no longer renames it. The method is rank-gated, and it no longer raises ValueError when the run is missing.
close is rank-gated	Only rank 0 finalizes the backends and saves the buffer.
version returns the run number	It returned the constant 1. It now returns the number prefix of the run name, or 0 for a user-supplied name.
A non-zero rank can fail to find the run	_get_latest_run_name waits for a run directory that was not there yet. It falls back to the newest existing run after a 1 s grace period, and it raises RuntimeError after 30 s. Every worker of a normal DDP launch pays the grace period and logs a warning, because the class cannot identify rank zero. Pass run_name to avoid both.

Bugs fixed

Bug	Effect
experiment retried the whole MLflow init on every access	Every log_metric ran set_experiment and start_run again. An unreachable server thus stalled the training loop. A failed init now backs off for 60 s.
Call sites evaluated self.experiment["mlflow"].log_metric eagerly	KeyError('mlflow') on the first log when mlflow was not importable.
name.rsplit("/", 1) unpacked two names	ValueError for an image name without a /, such as log_image("loss_curve", …).
close() only spilled the MLflow buffer	The TensorBoard writer never flushed and the WandB run never finished.
json.dump had no default	A np.float32 metric value crashed the fallback and lost the whole buffer.
save_logs_locally	It never cleared the buffer, named images {idx}.png, assumed 3 channels, and copied artifacts through read_bytes().
_get_latest_run_name()	IndexError on an empty directory. The time.sleep(1) DDP hotfix was not a synchronization primitive.
isnumeric() guarded int()	A directory named ½-baseline or ²-run crashed the constructor. isdecimal() matches what int accepts.
MLFLOW_HTTP_REQUEST_MAX_RETRIES was overwritten	The tracker now uses setdefault, and only when MLflow is enabled.
WandB logged through the module	Global state broke with two trackers in one process. Logging now goes through the Run that init returns.
artifact.save()	Deprecated, and it left the artifact unlinked. run.log_artifact replaces it.
np.array2string truncated	TensorBoard received an abbreviated matrix above 1000 elements.
add_hparams received any value	TensorBoard accepts only scalar hyperparameters. The tracker now converts every other value, None included, to a string.
project_name was set to None	MLflow init cleared it when project_id was given. The dead artifacts_dir attribute is also gone.

New behaviour

• MLflow dispatch works by call name. The old bound-method comparison had no
  else, so an unknown function silently discarded data.
• A failed MLflow call goes into a bounded buffer of 500 calls, of which 50 can
  be images. The tracker replays the buffer in order.
• The tracker retries an unreachable server only once a minute. A failed init
  and a failed replay both arm that backoff, so an outage cannot stall the
  training loop.
• A full buffer drops the oldest call and warns once for each outage.
  Hyperparameters and artifacts go last, because no later call repeats them.
  They still give way to the call that has just arrived.
• MLflow can reject one call and accept the next. The tracker sends the call
  behind a failing one to tell the two cases apart, and it drops only a
  rejected call. Two rejected calls in a row still look like an outage, and
  the tracker then keeps the whole buffer.
• The tracker copies a buffered artifact at once, because callers delete the
  file after they hand it over. Each copy gets its own artifacts/<n>/, so
  two artifacts that share a file name stay apart.
• close(status) is idempotent. It reports a backend that fails to shut down
  instead of raising.
• The module carries no Any and no type suppression. experiment maps each
  backend to its own handle type, and the buffered calls and the local fallback
  carry the types that occur.
• The tracker ignores an MLflow call that arrives after close(). MLflow would
  open a fresh run for it, and no later close() would end that run. The
  tracker warns instead.
• LuxonisTracker is a context manager. It marks the run failed if the block
  raises.
• The local fallback appends to local_logs.json, so a second save keeps the
  records of the first.

Dependencies & Potential Impact

• The tracker extra now needs opencv-python~=4.10. The module imports cv2
  at import time, so pip install luxonis-ml[tracker] failed before.
• Two extras are new: tensorboard (tensorboard, torch) and wandb.
  Each backend SDK was undeclared before, so a user had to know to install it.
  all pulls both in. The tracker extra stays free of them, because a run
  only needs the backends it turns on, and luxonis_ml.tracker still imports
  with none of them present.
• The tracker no longer imports LuxonisFileSystem.
• luxonis-train reads experiment["mlflow"] and calls close() in
  LuxonisTrackerPL._finalize. That code needs a follow-up. See the note below.

Downstream note for luxonis-train

LuxonisTrackerPL._finalize duplicates what close() now does. Two problems
appear once luxonis-train moves off the pinned luxonis-ml==0.9.1:

1. _finalize calls self.close() from inside its MLflow branch and does not
   forward the status. close() then finishes the WandB run with exit 0, and
   the later finish(1) cannot change it.
2. _finalize reads self.experiment["mlflow"] before it calls close(). The
   new _init_mlflow stores that handle only after start_run succeeds, so an
   unreachable server raises KeyError.
3. core.py uploads the ONNX file, the archive, the log, and the config after
   _finalize has run. close() now ends the MLflow run, and the tracker
   refuses a call that arrives after close(). Those artifacts therefore reach
   no run, and the tracker warns for each one.

Drop _finalize in a follow-up, and upload the artifacts before close().

Deployment Plan

None / not applicable. The change is a library change with no rollout step.

Testing & Validation

tests/test_tracker/ holds 67 tests. They reach 100 % statement coverage of
luxonis_ml/tracker.

• test_tracker.py runs against monkeypatched TensorBoard, WandB, and MLflow
  doubles. The doubles record every call and can be told to fail.
• test_mlflow_integration.py starts a real mlflow server subprocess. It uses
  an sqlite backend, a local artifact root, and a free port. It checks the
  params, the metrics, the image and artifact paths, and the run status through
  MlflowClient. A second test kills the server mid-run. It then asserts that
  the buffered logs reach local_logs.json and that close() does not raise.
  Both tests skip if the server does not start.

AI Usage

Assisted-by: Claude:claude-opus-5

Submitted code was reviewed by a human:

The author is taking the responsibility for the contribution:

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: f246fc35-e82d-49c6-90c8-66341ca7a145

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

❌ Patch coverage is 99.80159% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.58%. Comparing base (1567e6b) to head (45b9ab0).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
tests/test_tracker/test_mlflow_integration.py	97.77%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #460      +/-   ##
==========================================
+ Coverage   95.27%   95.58%   +0.30%     
==========================================
  Files         181      186       +5     
  Lines       15015    16106    +1091     
==========================================
+ Hits        14306    15395    +1089     
- Misses        709      711       +2     

Flag	Coverage Δ
pytest-ubuntu-latest	95.58% <99.80%> (+0.30%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.