Skip to content

v1.19.7 — Hardening

Choose a tag to compare

@elRadix elRadix released this 04 Nov 11:57
· 118 commits to main since this release
e4aaa8d
Daikin Cycle ML

🔧 v1.19.7 — Hardening: Trustworthy Alerts, Accurate Cycles

This is a hardening release. No new integrations to configure, no breaking option changes — just the kind of fixes that make the numbers on your dashboard tell the truth. Fourteen commits, seven bug families resolved, and one new diagnostic alert that quietly watches for a pattern the integration previously missed.

If you've ever seen a daily total above 24 hours, a 6-hour gap in a 24/7 heat pump run, or a short-cycle alert when you knew the compressor wasn't running — this release is for you.

🔍 Understanding the Issue

During a full end-to-end audit of Prod 1's database, three independent bugs turned out to produce the same symptom: numbers that don't add up. Each one alone damages COP, SPF, stooklijn-advies, and the ML baseline. Together they compound.

  • Cross-midnight cycles were counted 100% on the end-day. A 16-hour heating run from Oct 26 13:31 → Oct 27 05:46 appeared entirely as Oct 27 = 16h heating, and Oct 26 = 0h. On real data this produced days with 28.81 hours of heating in a 24-hour day — physically impossible.
  • Long low-modulation runs were completely invisible. On Nov 2 16:55–22:57 (6h02m), the compressor was actively heating — 69 COP samples, COP 5–7, flow 17 l/min, LWT climbing 28°C → 35°C → 31°C — yet the cycle detector recorded zero cycles in that window. Reason: INV frequency (rps) reports "Invalid" during low modulation after a defrost, and the detector only accepted RPS or a live-power fallback.
  • Two detection trackers could run at once. During BUH / defrost transitions, two open cycle trackers occasionally produced overlapping cycles with the same end_ts (e.g. id=175 6h and id=176 2h, both ending at 07:11). Every aggregation double-counted them.
  • Stale cycles from before a HA restart could fire alerts. After a reinstall or DB delete, HA hydrates the last completed cycle from .storage. If it had a small off_s stored, short_off fired even though sensor.cycles_today showed 0. Reported as #259, and directly triggered this whole audit.

✅ What We Changed

Alert correctness — 6 alert bugs aligned across both layers

  • short_off no longer fires on transient predecessors. A 30–120 second pump-only LWT sampling pulse before a real cycle no longer counts as a "short off".
  • short_off requires at least one short run in the pair. A 264-min → 60s → 84-min sequence is normal hysteresis restart, not short-cycling. Fires only when the run before or after is itself short.
  • Coordinator mirrors binary_sensor semantics. The dispatch path had drifted from the entity layer since v1.12.0 — three of these bugs were only fixed in one layer. All now match.
  • pendulum_hourly counts heating-only. Cooling and defrost no longer inflate the count in the coordinator path (matching the fix already in the entity layer since v1.19.6).
  • Corrupt is_transient="0" no longer disables short_run. Numeric parsing instead of a truthy check — a string "0" from a manual SQL edit or old backup is correctly treated as non-transient.
  • Stale cycles are ignored by short_run + short_off (#259). A cycle ending more than 1 hour ago is not treated as a live run or off-time. Fixes the false alerts after a HA restart that hydrates from a previous session.

Cycle-duration accuracy — three root causes fixed

  • Pro-rata cross-midnight attribution (#260). daily_summary now splits a cycle's duration across the days it actually ran. A 16h run from Oct 26 13:31 → Oct 27 05:46 reports ~10.5h on Oct 26 + ~5.75h on Oct 27, not 16h/0h.
  • thermal_kw_avg sample floor (#12). A 30s cycle with only 1 flow sample used to report its peak value as the average (33.79 kW observed on Prod 1 — physically impossible but under the magnitude cap). Cycles with fewer than 3 samples now store NULL.
  • Compressor detected via flow + COP (#261). When INV frequency (rps) reports "Invalid" — common during low modulation after a defrost — the compressor is still detected if flow > 5 L/min AND COP > 1.5 (better than resistive heat). Guarded against defrost-recovery (COP < 1.5) and pump-sampling-only (flow < 5). On real prod data this captures 68 of 69 missed samples in the Nov 2 gap.

New diagnostic alert — short_cycling_pattern

  • Structural pendeling detection. New binary_sensor.daikin_cycle_ml_short_cycling_pattern fires when 3+ short runs occur within a rolling 60-minute window. Distinct from single-cycle short_run: catches the behavioural pattern that single-cycle alerts miss.
  • Five diagnostic attributes so dashboards retain context even after the sensor returns to OFF: window_min, min_cycles_threshold, short_run_threshold_min, pattern_count_60min, recent_short_runs (last 10 short runs with timestamps and durations — newest first). Correlate past patterns with weather, LWT, or mode changes.
  • Native notification fires automatically. No user automation required. Dedup window: 60 minutes (same as pendulum_hourly). Group: short_cycle (bundles with short_run and short_off for grouped notifications).
  • Two new OptionsFlow fields: short_cycling_window_min (5–360, default 60) and short_cycling_min_cycles (1–20, default 3). Present in both setup wizard and options flow.

Silent bugs fixed that were hidden in plain sight

  • alerts DB table was dead code (#14). The alerts table existed since the very first commit but had zero call sites — nothing was ever inserted. Now db.async_insert_alert is wired into _async_dispatch_alerts. Existing rows stay empty on Prod 1/2/3; new alerts land immediately after upgrade.
  • Migration v27 — historical cleanup (#260, #261). An idempotent one-time migration on first HA restart after upgrade: finds overlapping root cycles produced by the dual-tracker bug, shortens the parent to child.start_ts, and rebuilds daily_summary for every affected day. On real Prod 1 data: 3 overlapping pairs fixed, 5 days rebuilt. Second run is a no-op (marker: data_migration_v27_completed).

Test coverage & CI hardening

  • +334 new tests across 14 new test files — E2E HVAC flow, edge cases, config-flow E2E, dispatch E2E, migration v27 with real fixture replay, cross-midnight attribution, compressor detection with 69-sample replay.
  • R596 gate now parses JUnit XML. Pytest 9.0.3 suppresses the N passed summary line on stdout, breaking the previous parser. The test-count check is now source-independent.
  • Ruff clean, mypy strict clean, markdown clean — full pre-push parity preserved.

📊 Impact for Your Heat Pump

What changes (visible):

  • sensor.daily_summary_recent will read lower for historical days. This is not a regression — it's the removal of double-counting. A day that used to show 19.4h because it absorbed the incoming half of a cross-midnight cycle now shows 18.3h, because the other 1.1h correctly belongs to the previous day.
  • sensor.cycles_duration_today on cold 24/7 days will now approach 21–23h heating + 1.5–2h DHW (was showing 16h or 28h depending on the bug).
  • Fewer short_off alerts — no more false positives on single cycles, transient predecessors, or stale hydrated cycles after a restart.
  • Fewer invisible 6h gaps — low-modulation heating is now detected via flow+COP.
  • New short_cycling_pattern binary sensor appears in HA and lights up during structural pendeling.
  • Notifications show actual counts — pendulum_daily message used to show 40 when the trigger counted 30 (transients excluded from trigger, not from message). Now consistent.
  • alerts table starts populating — after upgrade, SELECT COUNT(*) FROM alerts will reflect real alert history. Useful for diagnostics and ML correlation.

What stays the same:

  • All 17 previous alerts continue — short_run, short_off, pendulum_*, dhw_pendulum, high_cycle_rate, ml_anomaly, setpoint_osc, cop_low, cop_degradation, cop_vs_datasheet_low, defrost_excessive, buh_excessive, source_stale, missing_attributes, db_retention_warning, kmeans_retrain_stale, energy_daily_anomaly, stooklijn_advies.
  • Your configured thresholds are unchanged — short_run_threshold_min, pendulum_cycles_per_hour, dhw_pendulum_cycles_per_hour, dedup windows, quiet hours.
  • All 50 sensors and 13 services are unchanged — no reconfiguration needed.
  • The 5 Lovelace cards continue to work — no dashboard edits required.
  • Your database schema is preserved — v27 is additive (marker-only + targeted UPDATEs), no data loss.

What runs once on upgrade:

  • Migration v27 runs in the first HA restart after updating. On a database with months of history, this takes a few seconds (batched writes, per-day rollup). You may see these log lines once:
    v27 overlap cleanup: 3 pairs fixed, 5 days rebuilt
    data_migration_v27_completed: completed=True
    

🧪 Tests & Verification

  • Red → Green discipline preserved. Every fix was proven to fail on the old code first, then proven to pass on the new code. No green tests without a corresponding red test.
  • Production data replay. Compressor detection validated against 68 of 69 real COP samples from the Nov 2 gap on Prod 1. v27 migration validated against the 193-cycle Prod 1 fixture — 3 overlapping pairs found and fixed, exactly as expected.
  • Cross-midnight attribution proven with three separate scenarios: 16.25h cycle → ~10.5h + ~5.75h; single-day 8h cycle stays 8h; mixed sum preserved.
  • 24/7 week tests with synthetic data: cold week (5°C avg) → 20–24h/day heating; mild week (14°C avg) → 10–12h/day when the compressor genuinely modulates.
  • Resilience tests — v27 completes even if async_rollup_day or _rebuild_daily_summary_for_affected_days raises (defensive except blocks, but never silent: _LOGGER.exception fires).
  • +334 new tests across 14 new files — E2E, edge cases, migration fixture replay, dispatch E2E, config-flow E2E.
  • 4492 tests passing (was 4158 in v1.19.6) — full suite green, 100% line + branch coverage preserved, zero regressions.
  • Ruff + Mypy Strict + Markdown CI: all four gates pass.
  • R596 hardened — test-count verification now reads JUnit XML instead of stdout, resilient to pytest version changes.

⚙️ Upgrade Instructions

  • HACS Users: The update appears automatically within ~10 minutes. Or go to HACS → Daikin Cycle ML → Update.
  • Manual Install: Replace the custom_components/daikin_cycle_ml/ folder and restart Home Assistant.
  • Migration runs once on first restart. You do not need to trigger it manually. If you have a large history, expect 3–10 seconds of one-time DB activity. During the migration the integration continues to report data.
  • Action Required: If you disabled pendulum_hourly, pendulum_daily, dhw_pendulum, or high_cycle_rate notifications as a workaround, you can safely re-enable them now. See v1.19.6 notes for details.
  • No Configuration Changes: Your existing options, thresholds, and entity IDs are fully compatible.

📂 Dashboard Cards

Did you know we have 5 Lovelace cards for this integration? The COP & Energie card gives you a live efficiency overview at a glance — useful for spotting the same cross-midnight and gap patterns this release addresses.

🔗 Links & Support

Documentation:

Support:

Issues related to this release:

  • #258 — v1.19.7 hardening tracking
  • #259 — Stale cycle false positives (fixed)
  • #260 — Cross-midnight misattribution (fixed)
  • #261 — RPS-invalid detection gap (fixed)

☕ Support the Project

Daikin Cycle ML is free, local-only, and open-source — no cloud, no accounts, no telemetry. Your heat pump data stays on your Home Assistant instance. If this integration has helped you reduce short cycling, understand your cycles, or improve your COP, consider buying me a coffee. It goes a long way in funding future features and bug fixes.