What the Hekaton?

Hekaton, the power behind SQL Server 2014’s In-Memory OLTP technology, is intended to make data operations run orders of magnitude faster on SQL Server. This works its magic partly by serving database workloads entirely from main memory, using memory-optimized table structures. It replaces the relational engine’s standard locking model with an optimistic concurrency model based on time-stamped row versions. Deeper down the Hekaton engine uses new, ‘latch free’ data structures.

So far, so good, but performance improvements on this scale require a compromise, and the compromise is that these aren’t tables as we understand them. For the database developer, these differences are painful because they involve sacrificing some very important bits of the relational model. Most importantly, Hekaton tables don’t currently support FOREIGN KEY constraints or CHECK constraints, and you can’t put the checks in triggers because there aren’t any DML triggers either. Constraints allow a relational designer to enforce relational integrity and data integrity. Without them, of course, ‘bad data’ can get into our Hekaton tables. There is no easy way of preventing it. For several classes of database and data, this is a show-stopper.

One may regard all these restrictions regretfully, seeing limited opportunity to try out Hekaton with current databases, but perhaps there is also a sudden glow of recognition. Isn’t this how we all originally imagined table variables were going to be, back in SQL 2005? And they have much the same restrictions.

Maybe, instead of pretending that a currently-designed database can be ‘Hekatonized’ with a few mouse clicks, we should redesign databases for SQL 2014 to replace table variables with Hekaton tables, exploiting this technology for fast intermediate processing, and for the most part forget, for now, the idea of trying to convert our base relational tables into Hekaton tables. Few database developers would be averse to having their working tables running an order of magnitude faster, as long as it didn’t compromise the integrity of the data in the base tables.