There was a sharp intake of breath from the audience when Adam Machanic declared the SQL Server optimizer to be essentially “stuck in 1997”. It was during his fascinating “Query Tuning Mastery: Manhandling Parallelism” session at the recent PASS SQL Summit. Paraphrasing somewhat, Adam (blog | @AdamMachanic) offered a convincing argument that the optimizer often delivers flawed plans based on assumptions that are no longer valid with today’s hardware.

In 1997, when Microsoft engineers re-designed the database engine for SQL Server 7.0, SQL Server got its initial implementation of a cost-based optimizer. Up to SQL Server 2000, the developer often had to deploy a steady stream of hints in SQL statements to combat the occasionally wilful plan choices made by the optimizer. However, with each successive release, the optimizer has evolved and improved in its decision-making. It is still prone to the occasional stumble when we tackle difficult problems, join large numbers of tables, perform complex aggregations, and so on, but for most of us, most of the time, the optimizer purrs along efficiently in the background.

Adam, however, challenged further any assumption that the current optimizer is competent at providing the most efficient plans for our more complex analytical queries, and in particular of offering up correctly parallelized plans.

He painted a picture of a present where complex analytical queries have become ever more prevalent; where disk IO is ever faster so that reads from disk come into buffer cache faster than ever; where the improving RAM-to-data ratio means that we have a better chance of finding our data in cache. Most importantly, we have more CPUs at our disposal than ever before. To get these queries to perform, we not only need to have the right indexes, but also to be able to split the data up into subsets and spread its processing evenly across all these available CPUs.

Improvements such as support for ColumnStore indexes are taking things in the right direction, but, unfortunately, deficiencies in the current Optimizer mean that SQL Server is yet to be able to exploit properly all those extra CPUs. Adam’s contention was that the current optimizer uses essentially the same costing model for many of its core operations as it did back in the days of SQL Server 7, based on assumptions that are no longer valid. One example he gave was a “slow disk” bias that may have been valid back in 1997 but certainly is not on modern disk systems. Essentially, the optimizer assesses the relative cost of serial versus parallel plans based on the assumption that there is no IO cost benefit from parallelization, only CPU. It assumes that a single request will saturate the IO channel, and so a query would not run any faster if we parallelized IO because the disk system simply wouldn’t be able to handle the extra pressure. As such, the optimizer often decides that a serial plan is lower cost, often in cases where a parallel plan would improve performance dramatically.

It was challenging and thought provoking stuff, as were his techniques for driving parallelism through query logic based on subsets of rows that define the “grain” of the query. I highly recommend you catch the session if you missed it.

I’m interested to hear though, when and how often people feel the force of the optimizer’s shortcomings. Barring mistakes, such as stale statistics, how often do you feel the Optimizer fails to find the plan you think it should, and what are the most common causes? Is it fighting to induce it toward parallelism? Combating unexpected plans, arising from table partitioning? Something altogether more prosaic?

Cheers,

Tony.