Deepthi Goguri has two posts for us. First up, we learn about queries spilling to disk:
During the build time, the memory is allocated based on the cardinality estimates and the estimates are based on the input size and the probe. The buckets are created in the memory and we place the rows in those respective buckets. If the grant is exceeded then some of the buckets will be send to the disk. As some of the buckets are already in memory and some in disk, the initial probe of the data using the inner set. The probe rows will be scanned if the row hash to the in memory bucket the match is done for those rows. If the rows match to the on disk bucket, the probe row will be return to the disk along with the outer side bucket. So, because of this we have more disk writes to tempdb at the probe time. This is a build side spill.
If this final part of the series, lets talk about how we balance these memory grants. If lot of your queries are requesting for more memory than they required, then that causes concurrency issues. Memory grants are allocated to a any query even before they are executed based on the cardinality estimates and based on the memory consuming iterators. Once the query memory is allocated to a query, that memory is only released once the query execution completes even if the query actually uses a part of the allocated query. If more memory is requested by the queries than they need, then we are wasting the memory. What happens if the queries only receive less memory grants than they need, then we there is a performance impact on the running queries.
Read on for ways to fix excessive or insufficient memory grant problems.