The mechanics of the WHERE clause

Question

I’ve been out in industry for over 10 years now doing DB work most of that time. Lately I’ve been doing some work that is starting to challenge my long held assumptions about the internal workings of SQL Server, specifically what exactly is happening when
I do joins and where clauses.
I was explaining to someone that when doing a join putting in a where clause limits the amount of records that are available to do the join on thereby making the query faster. My thought was if a table has 100MM records but only 50MM of those meet the WHERE
criteria then the JOIN clause wouldn’t try matching the 50MM records that don’t meet the WHERE clause.
I did some work that afternoon and the empirical results didn’t match the expected outcome of a query I had written in terms of execution time. I don’t think my understanding of what goes on during a JOIN is correct. Can someone educate me?

Answer 1

Here is how a SELECT works in SQL ... at least in theory.  Real products will optimize things, but the code has to produce the same results. 
 a) Effectively materialize the CTEs in the optional WITH clause. CTE's come into existence in the order they are declared so only backward references are
alllowed. A CTE can be recursive. Think of them as VIEWs that exist only in the scope of the query. In practice, if they are used once then they are implemented as an in-line macro.
 b) Start in the FROM clause and build a working table from all of the joins, unions, intersections, and whatever other table constructors are there.  The <table expression> AS <correlation name> option allows you give a name to this working
table which you then have to use for the rest of the containing query.  Ther are UNIONB, INTERSECT and EXCEPT set construtors, LATERAL tables, table-valued funcitosn and all kinds of things happening in here. 
 c) Go to the WHERE clause and remove rows that do not pass criteria; that is, that do not test to TRUE (i.e. reject UNKNOWN and FALSE).  The WHERE clause is applied to the working set in the FROM clause.  
 d) Go to the optional GROUP BY clause, partiton the original table into groups and reduce each grouping to a *single* row, replacing the original working table with the new grouped table. The rows of a grouped table must be only group characteristics:
(1) a grouping column (2) a statistic about the group (i.e. aggregate functions) (3) a function or constant(4) an expression made up of only those three items.  The original table no longer exists and you cannot reference anything in it (this was an error
in early Sybase products).  
 e) Go to the optional HAVING clause and apply it against the grouped working table; if there was no GROUP BY clause, treat the entire table as one group. 
 f) Go to the SELECT clause and construct the expressions in the list. This means that the scalar subqueries, function calls and expressions in the SELECT are done after all the other clauses are done.  The AS operator can also give names to expressions
in the SELECT list.  These new names come into existence all at once, but after the WHERE clause, GROUP BY clause and HAVING clause have been executed; you cannot use them in the SELECT list or the WHERE clause for that reason. 
If there is a SELECT DISTINCT, then redundant duplicate rows are removed.  For purposes of defining a duplicate row, NULLs are treated as matching (just like in the GROUP BY).  
 g) Nested query expressions follow the usual scoping rules you would expect from a block structured language like C, Pascal, Algol, etc.  Namely, the innermost queries can reference columns and tables in the queries in which they are contained.  
 h) The ORDER BY clause is part of a cursor, not a query. The result set is passed to the cursor, which can only see the names in the SELECT clause list, and the sorting is done there.  The ORDER BY clause cannot have expression in it, or references
to other columns because the result set has been converted into a sequential file structure and that is what is being sorted.  
As you can see, things happen "all at once" in SQL, not "from left to right" as they would in a sequential file/procedural language model. In those languages, these two statements produce different results:
  READ (a, b, c) FROM File_X;
  READ (c, a, b) FROM File_X;
while these two statements return the same data:
SELECT a, b, c FROM Table_X;
SELECT c, a, b FROM Table_X;
Think about what a confused mess this statement is in the SQL model.
SELECT f(c2) AS c1, f(c1) AS c2 FROM Foobar;
That is why such nonsense is illegal syntax.
--CELKO-- Books in Celko Series for Morgan-Kaufmann Publishing: Analytics and OLAP in SQL / Data and Databases: Concepts in Practice Data / Measurements and Standards in SQL SQL for Smarties / SQL Programming Style / SQL Puzzles and Answers / Thinking
in Sets / Trees and Hierarchies in SQL