Itzik Ben-Gan continues a series on table expressions:
Last month I explained and demonstrated that CTEs get unnested, whereas temporary tables and table variables actually persist data. I provided recommendations in terms of when it makes sense to use CTEs versus when it makes sense to use temporary objects from a query performance standpoint. But there’s another important aspect of CTE optimization, or physical processing, to consider beyond the solution’s performance—how multiple references to the CTE from an outer query are handled. It’s important to realize that if you have an outer query with multiple references to the same CTE, each gets unnested separately. If you have nondeterministic calculations in the CTE’s inner query, those calculations can have different results in the different references.
Say for instance that you invoke the SYSDATETIME function in a CTE’s inner query, creating a result column called dt. Generally, assuming no change in the inputs, a built-in function is evaluated once per query and reference, irrespective of the number of rows involved. If you refer to the CTE only once from an outer query, but interact with the dt column multiple times, all references are supposed to represent the same function evaluation and return the same values. However, if you refer to the CTE multiple times in the outer query, be it with multiple subqueries referring to the CTE or a join between multiple instances of the same CTE (say aliased as C1 and C2), the references to C1.dt and C2.dt represent different evaluations of the underlying expression and could result in different values.
Definitely worth the read.