Plan değişimi P457

(@now datetime)WITH p AS (SELECT PlanHash, PlanXml, ROW_NUMBER() OVER (PARTITION BY PlanHash ORDER BY plan_id) AS rn FROM #px)
        INSERT perf.QdsPlanXml (PlanHash, PlanXmlCompressed, FirstSeenUtc, LastSeenUtc)
        SELECT PlanHash, COMPRESS(PlanXml), @now, @now FROM p
        WHERE rn = 1 AND NOT EXISTS (SELECT 1 FROM perf.QdsPlanXml AS e WHERE e.PlanHash = p.PlanHash)

Plan change case #P457 | DB: DB-12 | Regression | Confidence: Low (review pending)
Query: (@now datetime)WITH p AS (SELECT PlanHash, PlanXml, ROW_NUMBER() OVER (PARTITION BY PlanHash ORDER BY plan_id) AS rn FROM #px) INSERT perf.QdsPlanXml (PlanHash, PlanXmlCompressed, FirstSeenUtc, LastSeenUtc) SELECT PlanHash, COMPRESS(PlanXml), @now, @now FROM p WHERE rn = ? AND NOT EXISTS (SELECT ? FROM perf.QdsPlanXml AS e WHERE e.PlanHash = p.PlanHash)
Change: 2026-10-02 12:00 UTC, plan 7890 → 7895, pattern: Alternating (4 switches)
Impact (Query Store, per execution): duration 11.8 → 127.8 ms (+987%); CPU 6.4 → 71.1 ms (+1013%); reads 126 → 2,335 (+1748%); rows 2.81 → 79.34 (+2727%)
Plan difference: no operator-level difference
Candidate cause: STATS — Statistic perf.QdsPlanXml.PK_QdsPlanXml was updated 2026-10-01 22:10 (sampled 38.8%); 2 updates in the window; the plans use this statistic.
Candidate cause: WORKLOAD — The plans see a different workload: rows per execution 2.81 → 79.34. The plans may have been compiled for different parameters or batch sizes.
Candidate cause: PARAM — The two plans alternate in the same period (4 switches): plan choice depends on the parameter value rather than an event (parameter sensitivity).
Counter-evidence: WORK_PROPORTIONAL — Rows per execution changed 28.27× and duration 10.87×; duration per row 4.1895 → 1.6114 ms: the duration difference comes from the amount of data processed, not the plan.
Counter-evidence: HIGH_VARIANCE — The old plan's duration standard deviation is 409% of the average: the average is not reliable.
Counter-evidence: HIGH_VARIANCE — The new plan's duration standard deviation is 415% of the average: the average is not reliable.
Counter-evidence: METADATA_GAP — Metadata collection started on 2026-10-01 and the change was on 2026-10-02: index, statistics and module events could not be evaluated (no data, not no events).
Counter-evidence: SAME_SHAPE — The two plans have the same operator tree (operators, tables, indexes): the plan did not change; the difference is only in the compiled parameter values and row estimates. The slowness may come from the amount of data this parameter value processes or from waits, not from plan choice.
Confidence rationale: Low: strong counter-evidence (WORK_PROPORTIONAL, SAME_SHAPE).
Note: this is a candidate cause, not proof of the root cause.