DBAtlas · Database operations atlas

Find plan regressions with Query Store (same query, different plans)

SQL ServerQuery Store

Command

USE [Musteri];
-- Queries with more than one plan in the last 24 hours, per-plan performance:
WITH plan_stats AS (
    SELECT p.query_id, p.plan_id, p.is_forced_plan,
           SUM(rs.count_executions) AS execs,
           SUM(rs.avg_duration * rs.count_executions) / SUM(rs.count_executions) / 1000.0     AS avg_ms,
           SUM(rs.avg_cpu_time * rs.count_executions) / SUM(rs.count_executions) / 1000.0     AS avg_cpu_ms,
           SUM(rs.avg_logical_io_reads * rs.count_executions) / SUM(rs.count_executions)      AS avg_reads,
           MAX(rs.last_execution_time) AS last_exec
    FROM   sys.query_store_plan p
      JOIN sys.query_store_runtime_stats rs ON rs.plan_id = p.plan_id
      JOIN sys.query_store_runtime_stats_interval i
        ON i.runtime_stats_interval_id = rs.runtime_stats_interval_id
    WHERE  i.start_time >= DATEADD(HOUR, -24, SYSUTCDATETIME())
    GROUP BY p.query_id, p.plan_id, p.is_forced_plan
)
SELECT ps.query_id, ps.plan_id, ps.is_forced_plan, ps.execs,
       CAST(ps.avg_ms AS DECIMAL(18,2))     AS avg_ms,
       CAST(ps.avg_cpu_ms AS DECIMAL(18,2)) AS avg_cpu_ms,
       CAST(ps.avg_reads AS BIGINT)         AS avg_reads,
       CAST(ps.avg_ms / NULLIF(MIN(ps.avg_ms) OVER (PARTITION BY ps.query_id), 0) AS DECIMAL(18,1)) AS x_slower_than_best,
       ps.last_exec,
       OBJECT_NAME(q.object_id)     AS module,
       LEFT(qt.query_sql_text, 120) AS query_text
FROM   plan_stats ps
  JOIN sys.query_store_query q       ON q.query_id = ps.query_id
  JOIN sys.query_store_query_text qt ON qt.query_text_id = q.query_text_id
WHERE  ps.query_id IN (SELECT query_id FROM plan_stats GROUP BY query_id HAVING COUNT(*) > 1)
ORDER BY ps.query_id, ps.avg_ms;

What it does

Finds queries that ran with more than one plan in the last 24 hours; puts each plan's execution count, weighted average duration/CPU/reads and how many times slower it is than the best plan side by side.

When to use it

When a query "was fast yesterday and slow today", after a statistics update, index change, compatibility level upgrade or patch.

Caution

Average duration also depends on data distribution; when comparing plans that ran with different parameter values, look at avg_reads too (parameter sniffing). Time windows are UTC. Results are statistically weak for plans with few executions. Read-only query; safe.

Open in DBAtlas →

Related