SQL Server:联合后的内部联接导致哈希匹配速度慢（聚合）

下面是一个使整个存储过程变慢的CTE：

select * 
from #finalResults
where intervalEnd is not null
union
select            
    two.startTime, 
    two.endTime,
    two.intervalEnd,
    one.barcodeID,
    one.id,
    one.pairId,
    one.bookingTypeID,
    one.cardID,
    one.factor,
    two.openIntervals,
    two.factorSumConcurrentJobs
from #finalResults as one
inner join #finalResults as two
    on  two.cardID = one.cardID
    and two.startTime > one.startTime
    and two.startTime < one.intervalEnd

---

我是否也需要在此基础上构建一个单独的统计数据？

我曾经遇到过这样的情况：UNION的查询速度比UNION的查询速度慢得多，而UNION的所有查询都有一个不同的结果。因此，虽然我无法解释错误的查询计划统计信息和索引是否正常，但我建议您尝试以下方法：

select distinct * from (
    select * 
    from #finalResults
    where intervalEnd is not null
    union all
    select            
        two.startTime, 
        two.endTime,
        two.intervalEnd,
        one.barcodeID,
        one.id,
        one.pairId,
        one.bookingTypeID,
        one.cardID,
        one.factor,
        two.openIntervals,
        two.factorSumConcurrentJobs
    from #finalResults as one
    inner join #finalResults as two
        on  two.cardID = one.cardID
        and two.startTime > one.startTime
        and two.startTime < one.intervalEnd
)

---

如果您为cardID列创建统计信息，它可能会有所改进。

我已经这样做了：它提供了与使用optionorder组从原始CTE中选择相同的执行计划，即它花费的时间与以前相同，它只是将其用于排序而不是哈希。是的，查看该查询计划，散列匹配只有一个输入-因此是联合删除重复项导致了问题，我看不到任何关于连接的提示。也许使用一个UNION ALL，您可以确保不使用Èexcept添加重复记录，因为这样可以避免对第一部分的580K行进行排序？它不是散列联接，而是散列匹配。这是联合运算符注释的重复删除，它只有一个输入

select distinct * from (
    select * 
    from #finalResults
    where intervalEnd is not null
    union all
    select            
        two.startTime, 
        two.endTime,
        two.intervalEnd,
        one.barcodeID,
        one.id,
        one.pairId,
        one.bookingTypeID,
        one.cardID,
        one.factor,
        two.openIntervals,
        two.factorSumConcurrentJobs
    from #finalResults as one
    inner join #finalResults as two
        on  two.cardID = one.cardID
        and two.startTime > one.startTime
        and two.startTime < one.intervalEnd
)