Row Goals And Semi Joins

Paul White continues his row goals series:

The remaining physical join type is nested loops, which comes in two flavours: regular (uncorrelated) nested loops and apply nested loops (sometimes also referred to as a correlated or lateral join).

Regular nested loops join is similar to hash and merge join in that the join predicate is evaluated at the join. As before, this means there is no value in setting a row goal on either input. The left (upper) input will always be fully consumed eventually, and the inner input has no way to determine which row(s) should be prioritized, since we cannot know if a row will join or not until the predicate is tested at the join.

By contrast, an apply nested loops join has one or more outer references (correlated parameters) at the join, with the join predicate pushed down the inner (lower) side of the join. This creates an opportunity for the useful application of a row goal. Recall that a semi join only requires us to check for the existence of a row on join input B that matches the current row on join input A (thinking just about nested loops join strategies now).

In other words, on each iteration of an apply, we can stop looking at input B as soon as the first match is found, using the pushed-down join predicate. This is exactly the sort of thing a row goal is good for: generating part of a plan optimized to return the first n matching rows quickly (where n = 1 here).

This has the depth and quality that you naturally expect from Paul, making it an immediate read.

Related Posts

Creating Graph Tables in SQL Server

Mala Mahadevan continues a series on graph tables in SQL Server: I have highlighted in red what SQL Server adds to the table – the two system columns – graph id, which is bigint, and node id, which is nvarchar and stores json, and the unique index to help with queries. We can also see […]

Read More

Batch Mode Normalization

Paul White digs into batch mode normalization and its consequences for performance: I mentioned in the introduction that not all eight-byte data types can fit in 64 bits. This fact is important because many columnstore and batch mode performance optimizations only work with data 64 bits in size. Aggregate pushdown is one of those things. There are […]

Read More


February 2018
« Jan Mar »