Last month, Phil Factor caused a furore amongst some MVPs with an article that attempted to offer simple advice to developers regarding the use of table variables, versus local and global temporary tables, in their code. Phil makes clear that the table variables do come with some fairly major limitations.no distribution statistics, no parallel query plans for queries that modify table variables.but goes on to suggest that for reasonably small-scale strategic uses, and with a bit of due care and testing, table variables are a "good thing". Not everyone shares his opinion; in fact, I imagine he was rather aghast to learn that there were those felt his article was akin to pulling the pin out of a grenade and tossing it into the database; table variables should be avoided in almost all cases, according to their advice, in favour of temp tables. In other words, a fairly major feature of SQL Server should be more-or-less 'off limits' to developers.

The problem with temp tables is that, because they are scoped either in the procedure or the connection, it is easy to allow them to hang around for too long, eating up precious memory and bulking up the shared tempdb database. Unless they are explicitly dropped, global temporary tables, and local temporary tables created within a connection rather than within a stored procedure, will persist until the connection is closed or, with connection pooling, until the connection is reused. It's also quite common with ASP.NET applications to have connection leaks, as Bill Vaughn explains in his chapter in the "SQL Server Deep Dives" book, meaning that the web page exits without closing the connection object, maybe due to an error condition. This will then hang around in the heap for what might be hours before picked up by the garbage collector.

Table variables are much safer in this regard, since they are batch-scoped and so are cleaned up automatically once the batch is complete, which also means that they are intuitive to use for the developer because they conform to scoping rules that are closer to those in procedural code. On the surface then, an ideal way to deal with issues related to tempdb memory hogging.

So why did Phil qualify his recommendation to use Table Variables? This is another of those cases where, like scalar UDFs and table-valued multi-statement UDFs, developers can sometimes get into trouble with a relatively benign-looking feature, due to way it's been implemented in SQL Server. Once again the biggest problem is how they are handled internally, by the SQL Server query optimizer, which can make very poor choices for JOIN orders and so on, in the absence of statistics, especially when joining to tables with highly-skewed data. The resulting execution plans can be horrible, as will be the resulting performance. If the JOIN is to a large table, that will hurt.

Ideally, Microsoft would simply fix this issue so that developers can't get burned in this way; they've been around since SQL Server 2000, so Microsoft has had a bit of time to get it right. As I commented in regard to UDFs, when developers discover issues like with such standard features, the database becomes an alien planet to them, where death lurks around each corner, and they continue to avoid these "killer" features years after the problems have been eventually resolved.

In the meantime, what is the right approach? Is it to say "hammers can kill, don't ever use hammers", or is it to try to explain, as Phil's article and follow-up blog post have tried to do, what the feature was intended for, why care must be applied in its use, and so enable developers to make properly-informed decisions, without requiring them to delve deep into the inner workings of SQL Server?

Cheers,
Tony.