Bound per-track history and shrink the completed-track archive - #25
Conversation
Two per-Tracker stores grew far past anything that reads them, and the server holds one Tracker per radar node with days of uptime, so both are multiplied by the fleet. `Track.history` was five plain lists appended on every processed frame, on the update path and the coast path alike, with no cap anywhere. A track that never satisfies should_delete -- a loitering aircraft, or a persistent clutter return that coasts and re-associates forever -- therefore accumulated one timestamp, one frame number, one numpy state copy, one detection dict and one status string per frame for its whole life. Measured at 450 B per entry: 1.55 MiB per track-hour at 1 Hz, and strictly monotonic. This is the one store here that produces sustained growth rather than a plateau. The five buffers become deque(maxlen=TRACK_HISTORY_MAX), default 600 -- 10 minutes at 1 Hz, and about 60x the deepest reader. Every reader wants only the tail. `_initiate_tracklets` fits the last 3 associated samples and only ever runs on TENTATIVE tracks, which die at N_WINDOW frames. `get_recent_detections` is called with n bounded by the tracker's detection_window, 20 by default. `_merge_tracks` reads states[-1] and states[0] across a 5 s window. The server backend reads no history attribute at all: it goes through get_recent_detections, and asks for at most N2_TRACK_HISTORY_MAX, which is 20. The cap is overridable per Track with config tracker.track_history_max, for a batch consumer that wants deeper history than a long-running server should hold. Nothing slices the buffers, so the deque swap needed three adaptations rather than a rewrite. `get_recent_detections` walked its two lists by descending index, which on a deque is O(n) per access and would have made the scan quadratic; it now walks reversed() iterators of the two buffers in step, which preserves the semantics exactly -- newest-first, skipping the None measurements that mark_missed appends for coasted frames, stopping at n, returned oldest-first. `to_dict` now materializes the timestamps and state_status buffers with list(), because json.dumps cannot serialize a deque and the CLI dumps that structure. The `.extend()` calls in `_merge_track_pair` need no change. That merge also extended four of the five buffers and not "frames", which left the buffers unequal in length. Nothing reads "frames" today, so this was invisible, but with a shared maxlen the misalignment is the kind that outlives its cause. It extends all five now. `completed_tracks` is a deque of whole retired Track objects, each with its history attached, and MAX_COMPLETED_TRACKS was 5000 -- about 149 MiB per node at the measured 30 KiB for a 40-frame track, filled by ordinary track churn within the hour and then never released. Nothing outside this library and its tests reads it: get_confirmed_tracks and completed_tracks have zero call sites in the server, which reads only tracker.tracks. The ceiling drops to 50. Capping the count rather than stripping history off retired tracks, because the history is what makes the archive worth keeping: to_dict serializes each confirmed track's full history, and that is the CLI's entire batch output. A dropped payload would leave the retention tests asserting over hollow objects. A consumer that genuinely wants a deeper archive passes max_completed_tracks to Tracker(), so the ceiling is a default rather than a limit. The bound is pinned by tests over a 5000-frame run rather than by the constant alone: the buffers stay at maxlen while n_frames runs past it, and get_recent_detections still returns the newest 20 associated samples oldest-first, contiguous, ending on the frame just processed. A separate case marks three frames missed and asserts the None measurements are skipped, since that reverse-scan semantic is what the server's freshness signal depends on. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
Claude finished @jehanazad's task in 6m 21s —— View job Code Review
Reviewed 1.
So every real deployment silently gets the new hardcoded default of 50 instead of 5000. Per the PR's own reasoning, this archive is what 2. New test file adds inline comments, violating this repo's no-comments rule ( Both 3. The new
Everything else checks out: the list→deque migration doesn't break any reader ( |
1 participant
Summary
Two stores in the tracker grew without bound in a long-running server holding one
Trackerper node (~60 pipelines, days of uptime):Track.history(five plain lists appended on every update and every coast step, ~1.5 MiB per track-hour at 1 Hz, worse for a persistent clutter track) andTracker.completed_tracks(deque(maxlen=5000)of wholeTrackobjects with full history per node, read by nothing in the server). This bounds both.Changes
Track.historybuffers arecollections.deque(maxlen=TRACK_HISTORY_MAX);TRACK_HISTORY_MAX = 600(10 min at 1 Hz), overridable per track via configtracker.track_history_maxfor batch consumers that want deep history. The deepest reader in the server asks for 20 samples.get_recent_detectionsrewritten as a reversedzipover measurements/timestamps (deque indexing is O(n); semantics unchanged: newest-first scan, skips theNonea missed frame appends, returns the newest n oldest-first).to_dictmaterialises the deques (json.dumpscannot serialise a deque).MAX_COMPLETED_TRACKS5000 → 50, with amax_completed_trackskwarg onTracker.__init__for consumers that want the deeper archive. Count-capping rather than stripping history keepsto_dict()/the CLI batch output meaningful._merge_track_pairnow extends all five buffers (it skippedframes, leaving lengths unequal).Measured
Track.history: 450 B per entry → 1.55 MiB per track-hour unbounded, now 264 KiB flat per track.completed_tracks: ~30 KiB per retired track → per-node ceiling 149 MiB → 1.5 MiB.Tests
New
tests/test_history_bounds.py(6): 5000-frame run keeps every buffer at the cap;get_recent_detections(20)still returns the newest 20 associated samples in order; coasted frames skipped; configurable cap; bounded archive with the escape hatch; retired tracks keep bounded history. Suite 108 passed; pre-commit clean.Consumer: offworldlabs/retina-server pin bump follows.
🤖 Generated with Claude Code