Plan değişimi P450

(@from_utc datetime2(7),@to_utc datetime2(7),@epoch datetime2(7),@bucket int)INSERT INTO dbo.instance_counters_1m
    (bucket, instance_id, metric, scope, avg_value, max_value, min_value, last_value, samples)
SELECT k.bucket,
       k.instance_id,
       k.metric,
       k.scope,
       AVG(k.value),
       MAX(k.value),
       MIN(k.value),
       -- SON DEGER PENCERENIN SONUNDAN OKUNUR, MAX'TAN DEGIL:
       -- kumulatif bir sayacta ikisi ayni cikar, dalgalanan bir
       -- sayacta ayrilir ve "son" sozcugu MAX'a kaydirilirsa
       -- kolon adiyla icerigi celisir.
       --
       -- SIRA NUMARASI TEK GECISTE HESAPLANIR. Onceki bicim
       -- kovanin son anini `CROSS APPLY` ile HER SATIR ICIN
       -- yeniden ariyordu; olculdu (2026-09-01): iki bin
       -- dakikalik bir yakalama turu 20 saniyeden 900 saniyeye
       -- kadar cikti ve motor kendi ilk dolumunu bitiremedi.
       -- Pencere islevi ayni cevabi tek taramada verir --
       -- p95'te de ayni cozum kullanildi.
       MAX(CASE WHEN k.rn = 1 THEN k.value END),
       COUNT_BIG(*)
FROM (
    SELECT DATEADD(SECOND, (DATEDIFF_BIG(SECOND, @epoch, c.ts) / @bucket) * @bucket, @epoch) AS bucket,
           c.instance_id,
           c.metric,
           c.scope,
           c.value,
           ROW_NUMBER() OVER (
               PARTITION BY
                   DATEADD(SECOND, (DATEDIFF_BIG(SECOND, @epoch, c.ts) / @bucket) * @bucket, @epoch),
                   c.instance_id, c.metric, c.scope
               ORDER BY c.ts DESC) AS rn
    FROM dbo.instance_counters AS c
    WHERE c.ts >= @from_utc AND c.ts < @to_utc
) AS k
GROUP BY k.bucket, k.instance_id, k.metric, k.scope

Plan change case #P450 | DB: DB-10 | Improvement | Confidence: Medium (review pending)
Query: (@from_utc datetime2(?),@to_utc datetime2(?),@epoch datetime2(?),@bucket int)INSERT INTO dbo.instance_counters_1m (bucket, instance_id, metric, scope, avg_value, max_value, min_value, last_value, samples) SELECT k.bucket, k.instance_id, k.metric, k.scope, AVG(k.value), MAX(k.value), MIN(k.value), MAX(CASE WHEN k.rn = ? THEN k.value END), COUNT_BIG(*) FROM ( SELECT DATEADD(SECOND, (DATEDIFF_BIG(SECOND, @epoch, c.ts) / @bucket) * @bucket, @epoch) AS bucket, c.instance_id, c.metric, c.scope, c.value, ROW_NUMBER() OVER ( PARTITION BY DATEADD(SECOND, (DATEDIFF_BIG(SECOND, @epoch, c.ts) / @bucket) * @bucket, @epoch), c.instance_id, c.metric, c.scope ORDER BY c.ts DESC) AS rn FROM dbo.instance_counters AS c WHERE c.ts >= @from_utc AND c.ts < @to_utc ) AS k GROUP BY k.bucket, k.instance_id, k.metric, k.scope
Change: 2026-09-28 10:00 UTC, plan 582 → 627, pattern: Transition
Impact (Query Store, per execution): duration 1935 → 1480.1 ms (-24%); CPU 183.1 → 74.4 ms (-59%); reads 12,107 → 6,397 (-47%); rows 995.36 → 996.53 (+0%)
Plan difference: no operator-level difference
Candidate cause: FORCED — The old plan was forced and the new plan is not: forcing was removed or failed.
Counter-evidence: HIGH_VARIANCE — The old plan's duration standard deviation is 794% of the average: the average is not reliable.
Counter-evidence: HIGH_VARIANCE — The new plan's duration standard deviation is 1688% of the average: the average is not reliable.
Counter-evidence: METADATA_GAP — Metadata collection started on 2026-09-27 and the change was on 2026-09-28: index, statistics and module events could not be evaluated (no data, not no events).
Confidence rationale: Medium: a single mechanical candidate (FORCED) but weak counter-evidence (HIGH_VARIANCE, METADATA_GAP).
Note: this is a candidate cause, not proof of the root cause.