SQL Server索引用法取决于列值

Question

I have a huge table (it has more than 40 milion records) and an index on one of the column. The column is a reference column - typeId.

I run into an issue that while filtering the data on that column using different column values, the index is sometimes being used and sometimes not.

Eg. simply running:

select * from LargeNumberOfItemsTable where typeid=4

makes SQL Server use index, but

select * from LargeNumberOfItemsTable where typeid=3

does not use index.

When the index is not being used, the query takes a long time, but when I force a hint that the index should be used, it takes just seconds to run it.

The strange thing is that statistics have been updated and the index has been rebuilt.

Does anyone have an idea, why such behaviour could occur? I'm looking for some hints where to look for the cause.

Answer 1

SQL Server chooses the plan based on estimated row counts. If the estimated row counts are accurate, the time taken for a full table scan versus a seek and key lookup will depend much on your hardware and available resources.

Make sure indexes were rebuilt with MAXDOP 1. Otherwise, the out-of-order extents will mitigate the benefits of read-ahead scans, causing the plan with the scan to perform more poorly when data are not cached.