Cut poller/backend CPU on small hosts: lazy entity builds, DR dedup, Whisper cap

[d3mocide]

Summary

Fixes the CPU regression observed after PR #117 — both cores pegged, load ~5.0 on a 2-core VPS. The hot paths were benchmarked with synthetic BEAST traffic; each change below addresses a measured cost, not a guess.

The regression

PR #117's dead reckoning made stale aircraft's lat/lon advance with wall-clock time. Those keys are in _entity_changed's compare set, so aircraft that previously went quiet when stale started re-publishing every second: Redis pub/sub → backend WS fan-out (per client, per worker) → DB write → frontend render. On top of that, several pre-existing per-frame and per-publish costs were paying for work that was thrown away.

Changes

Poller — ADS-B pipeline

• Lazy entity building (beast_decoder.py, adsb.py): the frame worker built a full entity dict (trail copy, comm-B snapshot, DR projection, ISO timestamp) on every decoded frame, but only published at ≤1/s per aircraft. ingest() is split into ingest_frame() (state-only) + entity_from_state(); the dict is now built only when the publish gate passes. ~22% cheaper per frame; ingest() remains as a compat wrapper so existing tests/callers are untouched. Best Mode arbitration semantics preserved ("beast seen" still recorded only for positioned aircraft).
• Snapshot rebuild moved to snapshot cadence (adsb.py): snapshot_entities() — an O(aircraft) full entity rebuild — ran every tick (1s) while its only consumer, the enriched snapshot publish, runs every 5s. It now runs once per snapshot tick, immediately before the publish (80% less of that work, and the data is strictly fresher). DR still advances through a total feed outage.
• DR republish suppression (bus.py): while both previous and current state are dead-reckoned, _entity_changed skips lat/lon/altitude — they're synthetic projections, and the frontend already projects DR tracks client-side (pvb.ts). Real telemetry changes (heading, speed, vertical rate, squawk, …) and DR on/off transitions (position_dr added to the compare keys) still publish immediately.
• Single sanitize per publish (bus.py, db.py): write_entity_observation re-ran sanitize_payload — a recursive deep-copy that walks all 150 trail points, ~216µs/call — on entities its only caller had already sanitized. A sanitized=True flag skips the redundant pass, roughly halving per-publish CPU.
• BEAST transport fast path (beast_transport.py): frames whose wire body contains no 0x1A escape byte (the overwhelmingly common case) are sliced out at C speed via bytearray.find instead of the per-byte Python loop. Escaped, incomplete, and garbage input still goes through the unchanged parse_frame slow path.

Transcription

• New WHISPER_CPU_THREADS setting (default 1) passed to WhisperModel, and the container's compose CPU limit reduced 2.0 → 1.0. Previously ctranslate2 grabbed every core whenever a P25 call transcribed, starving the poller/backend/map on a 2-core host. Base/int8 on one thread still transcribes short P25 clips faster than realtime; raise the env var on bigger hosts.

Measurements (synthetic 50/50 short/long frame mix)

Path	Before	After
Transport frame parse	3.99 µs/frame	1.09 µs/frame (3.7×)
Decoder per-frame (worker path)	20.5 µs	16.8 µs (~22% incl. skipped entity builds)
Per-publish sanitize	2 × 216 µs	1 × 216 µs
DR'd stale aircraft publish rate	1/s each	only on real telemetry change

Tests

• New poller/tests/test_beast_transport.py (11 tests): fast path, escaped MLAT/signal/message bytes, garbage resync, incomplete buffers, 0x31 skipping — the parser previously had no tests.
• 5 new DR-dedup tests in test_bus.py.
• Full poller suite: 171 passed. tsc --noEmit clean, docker compose config clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_017n1uFNfMESakg9E9qoMDrh

---

Generated by Claude Code