I understand that the WITH RECOMPILE option forces the optimizer to rebuild the query plan for stored procs but when would you want that to happen?
What are some rules of thumb on when to use the WITH RECOMPILE option and when not to?
What's the effective overhead associated with just putting it on every sproc?
The most common use is when you might have a dynamic WHERE clause in a procedure...you wouldn't want that particular query plan to get compiled and saved for subsequent executions because it very well might not be the exact same clause the next time the procedure is called.
It should only be used when testing with reprentative data and context demonstrate that doing without produces invalid query plans (whatever the possible reasons might be). Don't assume beforehand (without testing) that an SP won't optimize properly.
Sole exception for manual invocation only (i.e. don't code it into the SP): When you know that you've substantially altered the character of the target tables. e.g. TRUNCATE, bulk loads, etc.
It's yet another opportunity for premature optimization.
Note: I have plenty of points. If a newby submits the same answer below, and you agree, upvote theirs.
Putting it on every stored procedure is NOT a good idea, because compiling a query plan is a relatively expensive operation and you will not see any benefit from the query plans being cached and re-used.
The case of a dynamic where clause built up inside a stored procedure can be handled using sp_executesql to execute the TSQL rather than adding WITH RECOMPILE to the stored procedure.
Another solution (SQL Server 2005 onwards) is to use hint with specific parameters using the OPTIMIZE FOR hint. This works well if the values in the rows are static.
SQL Server 2008 has introduced a little known feature called "OPTIMIZE FOR UNKNOWN":
This hint directs the query optimizer to use the standard algorithms it has always used if no parameters values had been passed to the query at all. In this case the optimizer will look at all available statistical data to reach a determination of what the values of the local variables used to generate the queryplan should be, instead of looking at the specific parameter values that were passed to the query by the application.
As others have said, you don't want to simply include WITH RECOMPILE in every stored proc as a matter of habit. By doing so, you'd be eliminating one of the primary benefits of stored procedures: the fact that it saves the query plan.
Why is that potentially a big deal? Computing a query plan is a lot more intensive than compiling regular procedural code. Because the syntax of a SQL statement only specifies what you want, and not (generally) how to get it, that allows the database a wide degree of flexibility when creating the physical plan (that is, the step-by-step instructions to actually gather and modify data). There are lots of "tricks" the database query pre-processor can do and choices it can make - what order to join the tables, which indexes to use, whether to apply WHERE clauses before or after joins, etc.
For a simple SELECT statement, it might not make a difference, but for any non-trivial query, the database is going to spend some serious time (measured in milliseconds, as opposed to the usual microseconds) to come up with an optimal plan. For really complex queries, it can't even guarantee an optimal plan, it has to just use heuristics to come up with a pretty good plan. So by forcing it to recompile every time, you're telling it that it has to go through that process over and over again, even if the plan it got before was perfectly good.
Depending on the vendor, there should be automatic triggers for recompiling query plans - for example, if the statistics on a table change significantly (like, the histogram of values in a certain column starts out evenly distributed by over time becomes highly skewed), then the DB should notice that and recompile the plan. But generally speaking, the implementers of a database are going to be smarter about that on the whole than you are.
As with anything performance related, don't take shots in the dark; figure out where the bottlenecks are that are costing 90% of your performance, and solve them first.