At "Last 7 days" the WPF Queries grids silently show ~4 days: Top Queries, Top Procedures and Query Store read raw tables dropped at 4 days, with no floor disclosure (get_query_store_top got one in #2364; the viewer and the query-stats MCP tools didn't)

[erikdarlingdata]

Problem

The WPF server tab's toolbar offers "Last 7 days". Three Queries sub-tabs read only raw tables, and all three are dropped at 4 days on a store with the rollups armed:

grid	table
Top Queries by Duration	query_stats
Top Procedures by Duration	procedure_stats
Query Store by Duration, and its slicer	query_store_stats

None of these reads routes to a rollup, and none reports how far back it actually reached. At 7 days the grid ranks about 4.2 days of work under a header and toolbar that say 7, and nothing on screen says so.

#2364 fixed exactly this for MCP get_query_store_top: it reads the window floor and returns effective_start / window_truncated. The WPF Query Store grid, the WPF Top Queries and Top Procedures grids, and MCP get_top_queries_by_cpu / get_top_procedures_by_cpu never got the same treatment. The WPF viewer's own Performance Trends tab does disclose (QueryTrendSeries.Truncated, rendered in the plot title), so one tab of one screen is honest and the grid beside it isn't.

It's also the expensive way to be wrong: the 7-day read scans and dedupes all 4 days of raw.

Measured

2026-09-25, read-only:

• Raw floors. Production SQL Server store B: query_stats and procedure_stats both start at 2026-09-21 00:00, 4.19 days back at 04:32Z. Store A (live Query Store), one busy server: query_store_stats starts at 2026-09-21 00:02, 918,803 rows in the "7-day" window.
• The WPF Query Store grid's read (QueryStoreTopSql) at 7 days, same server: 4,078 ms, 162.9 k buffers (28 k read). The window sort spilled: external merge of 100 + 121 + 203 MB, 53 k temp blocks written, at that store's current work_mem. At 24 h the same read is 1,090 ms with no spill.

Where (origin/dev)

• Darling/PerformanceMonitor.Darling.Viewer/ViewerServerTab.Queries.cs: LoadTopQueriesAsync ~:172, LoadTopProceduresAsync ~:181, LoadQueryStoreAsync ~:190. None reads a floor or renders one.
• Darling/PerformanceMonitor.Darling.Viewer/ViewerDataService.QueryStats.cs TopQueriesSql ~:149, ViewerDataService.ProcedureStats.cs TopProceduresSql ~:88, ViewerDataService.QueryStore.cs QueryStoreTopSql ~:123 and QueryStoreSlicerSql ~:595.
• The disclosure that exists: Darling/PerformanceMonitor.Darling.Service/Mcp/DarlingMcpDataTools.cs get_query_store_top ~:796–880, and DarlingDataReader.QueryStoreWindowFloorSql ~:1323 (a bounded MIN(collection_time), a one-chunk probe).
• The tools without it: DarlingMcpDataTools.cs get_top_queries_by_cpu ~:515/:560 and get_top_procedures_by_cpu ~:671/:690, both over raw DarlingDataReader.TopQueriesSql / procedure reads.

Fix shape

• Disclose, in every grid and tool that reads a raw-only table:
  • read the window floor (the get_query_store_top reports a window it cannot serve: raw query_store_stats is dropped at 4 days #2364 statement, per table);
  • show "showing since " in the grid header when the floor is more than the get_query_store_top reports a window it cannot serve: raw query_store_stats is dropped at 4 days #2364 slack (90 min) after the requested start;
  • in the MCP tools, return effective_start / effective_hours_back / window_truncated like get_query_store_top.
• And/or route. query_stats and procedure_stats have query-grain hourly rollups (query_stats_hourly, procedure_stats_hourly) that the Performance Trends and FinOps reads already stitch through RollupCoverage. A top-N by total over a window past raw retention can be served from them, the same way the trend tier ladder does it. query_store_stats has no rollup that carries query_id/plan_id (get_query_store_top reports a window it cannot serve: raw query_store_stats is dropped at 4 days #2364's note), so its only honest option is the disclosure.
• Pins:
  • a live test on a seeded store whose raw table is shorter than the window: the WPF grid's header text, and each MCP tool's window_truncated / effective_start, name the floor;
  • a source pin that every Queries-tab grid read passes through the shared floor helper.

[erikdarlingdata]

Ruled: how this gets built

Three pieces, in this order:

1. Darling disclosure (now). One shared floor helper for query_stats, procedure_stats and query_store_stats, built from get_query_store_top reports a window it cannot serve: raw query_store_stats is dropped at 4 days #2364's QueryStoreWindowFloorSql. It must stay a one-chunk probe. The PR says "Part of At "Last 7 days" the WPF Queries grids silently show ~4 days: Top Queries, Top Procedures and Query Store read raw tables dropped at 4 days, with no floor disclosure (get_query_store_top got one in #2364; the viewer and the query-stats MCP tools didn't) #4231".
   • get_top_queries_by_cpu and get_top_procedures_by_cpu add effective_start, effective_hours_back and window_truncated, with the same names and meaning as get_query_store_top. Existing fields and defaults stay as they are.
   • The WPF Top Queries, Top Procedures and Query Store grids show "Showing since " in the header when the window is cut short. The web viewer's pages for the same tools show the same note.
   • The threshold is get_query_store_top reports a window it cannot serve: raw query_store_stats is dropped at 4 days #2364's 90-minute slack, reusing its constant.
2. Lite disclosure (next). Lite keeps 30 days of raw query data by default. A user can lower that, and a new install has less history, so the same silent cut can happen there. Same rules, Lite's own grids and tools.
3. Routing the top-N reads to the hourly rollups (after Legacy trio off the refresh grid (#3653) #4186 merges). This builds on RollupCoverage, and Legacy trio off the refresh grid (#3653) #4186 changes it. The disclosure ships first either way, and it is still needed for query_store_stats, which has no rollup that keeps query_id.

\U0001F916 Generated with Claude Code

https://claude.ai/code/session_01TszxYhJJbTEh4LrZ56NYo3

[erikdarlingdata]

Status: #4278 (the Darling half) merged to dev (f0370dc). Left: #4279 (the Lite half, next; its last step updates the shared tool descriptions in both products), and stage 3 (routing the top-N reads to the hourly rollups) after #4186.