FinOps Lite top-consumer and idle-check reads use one statement each (part of #4227)

[erikdarlingdata]

Part of #4227.

Why

Lane V3b confirmed by inspection that Lite's FinOps tab carries the same two redundant-read patterns #4240 fixed in the Darling viewer. The coordinator also added a third item: V3a's PR #4241 folded Darling's Server Inventory idle-database check from one statement per server into one fleet statement. V3b measured Lite's twin too.

This PR measures all three patterns on a seeded DuckDB before changing anything and ports only the ones with a real, repeatable win. DuckDB is columnar and in-process. A second read over data DuckDB already cached can cost next to nothing. That differs from Postgres/TimescaleDB, where the original #4227/#4240/#4241 wins came from avoiding a second pass over a large, disk-resident hypertable. That difference held for one of the three patterns here.

What I found and measured

Seeded one DuckDB (own isolated instance, not the shared test fixture) with 50 servers and 15,000+ rows across cpu/memory/grant/size/query_stats. One focus server had a heavier load: 4,800 query_stats rows, 960 file_io_stats rows, and 4,500 index_object_stats rows over 30 daily snapshots. Timed the current code against a merged form, 5 iterations each, 3 full runs:

pattern	old	new	verdict
Utilization top-consumers (by total + by average)	132-220 ms/pair (2 statements, 2 connection-opens)	60-112 ms/pair (1 statement, 1 connection-open)	ported: consistent ~2x across all 3 runs
Storage Growth database drill (bounds computed once per statement, twice total)	97-209 ms/call	105-137 ms/call	not ported: direction flipped between runs (new was slower once, faster once). No reliable win
Server Inventory idle-database check (N per-server calls vs. 1 fleet statement)	128-185 ms wall-clock (50 servers, parallel Task.WhenAll, matches production)	85-99 ms	ported: consistent ~2x across all 3 runs

The Utilization and Server Inventory wins both trace mostly to halving or fleet-collapsing the OpenConnectionAsync/read-lock-acquisition cost. Each of those methods opens its own DuckDB connection today. The Storage Growth pattern's redundant work is a cheap MAX/MIN over an already-indexed collection_time column, computed within a single connection either way. There was nothing real to save.

"Lite's top server count"

Not documented anywhere I found (no cap in ServerManager, no stated limit in Lite/README.md). This PR uses 50, matching the production 43-server fleet #4227 used as a round-number ceiling.

Continuous aggregates

Lite has none (no TimescaleDB, no rollup views). The rollup/watermark routing #4241 added for the idle check does not apply here. The merged fleet statement reads raw tables only, the same as the per-server loop it replaces.

What changes

• Lite/Services/LocalDataService.FinOps.Workload.cs: GetTopResourceConsumersByTotalAsync and GetTopResourceConsumersByAvgAsync become one GetTopResourceConsumersAsync, returning (List<TopResourceConsumerRow> ByTotal, List<TopResourceConsumerRow> ByAvg) from one statement. Each grid keeps its own old filter. ByTotal drops a NULL database name. ByAvg drops a zero-execution database. Both filters now apply client-side from the same row set.
• Lite/Services/LocalDataService.FinOps.ServerProperties.cs: GetServerMetricsAsync(int serverId) becomes GetServerMetricsAsync(), returning Dictionary<int, ServerMetricsRow> for every server in one statement. A server with no collected rows still gets an entry (all fields null), matching the old per-server statement's anchor row.
• Lite/Controls/FinOpsTab.xaml.cs: LoadServerInventoryAsync calls the fleet statement once before the per-server fan-out. The fan-out still runs per server for the LIVE SQL query (a real network call per monitored server). Only the DuckDB overlay moved. LoadUtilizationAsync calls the merged consumer read once.
• No change to LocalDataService.FinOps.IndexObjects.cs (Storage Growth). Measured, not ported. Numbers are in the table above.

Test plan

• Lite.Tests/FinOpsFleetReadParityTests.cs (new): pins that the merged Utilization read keeps each old grid's NULL-name/zero-execution filter. It also pins that the fleet statement covers every server (a server with zero rows gets a slot) and flags an idle database. Both tests pass.
• FinOpsVerdictSourcePinTests (existing [BUG] FinOps provisioning status reports UNDER_PROVISIONED for every healthy server: the memory ratio compares Total to Target Server Memory, which converge at 1.0 #2246 source pin) still passes against the rewritten ServerProperties.cs. The pressure-signal columns and the shared ProvisioningVerdict.Evaluate call are still there in the merged SQL/C#.
• dotnet build Lite.Tests/Lite.Tests.csproj: 0 warnings, 0 errors, both before and after merging origin/dev (fast-forward, no conflicts, touched only Darling files).
• Full Lite.Tests suite: not run. This lane hit its context checkpoint after the targeted runs above. The coordinator must run a full-suite pass before merging.

CHANGELOG entry

SECTION: Fixed
ENTRY:

• FinOps tab loads faster on Lite ([FinOps Lite top-consumer and idle-check reads use one statement each (part of #4227) #4243]). The Utilization sub-tab's two top-consumer grids each scanned the same query and I/O data separately. They now share one read. Server Inventory's idle-database check ran once per monitored server. It now runs once for the whole fleet. Both changes are internal. The numbers shown do not change.
  REF:
  [FinOps Lite top-consumer and idle-check reads use one statement each (part of #4227) #4243]: FinOps Lite top-consumer and idle-check reads use one statement each (part of #4227) #4243

Generated with Claude Code

https://claude.ai/code/session_01TszxYhJJbTEh4LrZ56NYo3