Repository navigation
Unify the collector-health ROW banding (NEVER_RUN/FAILING/STALE/WARNING/HEALTHY) — currently defined in three places #1573
Description
Surfaced during the #1562 data-layer build. The per-server CARD banding (CPU/threads/memory/blocking/deadlocks/collectors + freshness) is now unified in PerformanceMonitor.Common/ServerHealthBands.cs (ServerHealthClassifier) and consumed by the WPF viewer, the web fleet reader, and the new get_fleet_overview MCP tool — one set of thresholds, drift closed.
The collector-health row banding is a different, larger classifier that still exists in three independent copies: Lite's collection-health surface, the Darling WPF viewer's, and the service's DarlingDataReader.CollectorHealth (NEVER_RUN / FAILING / STALE / WARNING / HEALTHY per collector row). They agree today; nothing pins them together, so they can drift exactly the way the card thresholds could before unification.
Scope of the fix: extract one row classifier (Common, brush/display-free), refactor all three consumers onto it, keep each surface's display formatting local, pin with a shared decision-table test. Entangled with WPF display-string formatting on both viewer sides, which is why it was consciously not folded into #1562's build.
Activity
erikdarlingdata commentedon Jul 18, 2026
Scope addition from the #1574 merge gate: the cross-app MCP inventory pin caught get_fleet_overview (born from the web dashboard's DarlingFleetReader) with no Lite twin — it is now on the KnownLiteMissingMcpTools ratchet with a pointer here. The Lite port belongs to this issue's unification work: a DuckDB fleet reader over the SAME shared ServerHealthClassifier in Common, which is also the moment to move Lite's own overview-card banding onto that classifier (currently Lite-side thresholds — the same pre-unification drift the Darling viewer had). Port the tool, remove the ratchet entry (it only shrinks).
Lite pin: ratchet-list get_fleet_overview (Darling-only until the #1573…
Unify collector-health ROW banding + fix daily-collector false-positi…
Merge pull request #1578 from erikdarlingdata/feature/1573-collector-…
erikdarlingdata commentedon Jul 18, 2026
Shipped to dev in #1578. Collector-health ROW banding is unified into one PerformanceMonitor.Common classifier (CollectorHealthClassifier) that Lite, the Darling viewer, and the service all delegate to — no more three-way drift. Thresholds are now frequency-relative: FAILING at max(24h, 2x cadence), STALE at max(4h, 1.5x cadence), so a healthy daily collector (index_object_stats, 1440-min) reads HEALTHY at 27h instead of FAILING, and the Overview 'N failing' count (which reuses this banding) stops showing phantom failures. Floors keep every frequent collector's banding byte-for-byte identical. 40 decision-table tests mirrored in both apps. NOTE: a 4th, SQL-side copy exists in the deprecated Dashboard (install/47) with different semantics and can't consume the C# classifier - out of scope, flagged in the PR.
Make the server stale band honor each collector's own cadence (erikda…