From NaN to Infinity…and Beyond!Published 13 April 2010 10:26 am
It is hard to believe that it was once possible to corrupt a SQL Server Database by storing perfectly normal data values into a table; but it is true. In SQL Server 2000 and before, one could inadvertently load invalid data values into certain data types via RPC calls or bulk insert methods rather than DML. In the particular case of the FLOAT data type, this meant that common ‘special values’ for this type, namely NaN (not-a-number) and +/- infinity, could be quite happily plugged into the database from an application and stored as ‘out-of-range’ values.
This was like a time-bomb. When one then tried to query this data; the values were unsupported and so data pages containing them were flagged as being corrupt. Any query that needed to read a column containing the special value could fail or return unpredictable results. Microsoft even had to issue a hotfix to deal with failures in the automatic recovery process, caused by the presence of these NaN values, which rendered the whole database inaccessible!
This problem is history for those of us on more current versions of SQL Server, but its ghost still haunts us. Recently, for example, a developer on Red Gate’s SQL Response team reported a strange problem when attempting to load historical monitoring data into a SQL Server 2005 database via the C# ADO.NET provider. The ratios used in some of their reporting calculations occasionally threw out NaN or infinity values, and the subsequent attempts to load these values resulted in a nasty error.
It turns out to be a different manifestation of the same problem. SQL Server 2005 still does not fully support the IEEE 754 standard for floating point numbers, in that the FLOAT data type still cannot handle NaN or infinity values. Instead, they just added validation checks that prevent the ‘invalid’ values from being loaded in the first place. For people migrating from SQL Server 2000 databases that contained out-of-range FLOAT (or DATETIME etc.) data, to SQL Server 2005, Microsoft have added to the latter’s version of the DBCC CHECKDB (or CHECKTABLE) command a DATA_PURITY clause. When enabled, this will seek out the corrupt data, but won’t fix it. You have to do this yourself in what can often be a slow, painful manual process.
Our development team, after a quizzical shrug of the shoulders, simply decided to represent NaN and infinity values as NULL, and move on, accepting the minor inconvenience of not being able to tell them apart. However, what of scientific, engineering and other applications that really would like the luxury of being able to both store and access these perfectly-reasonable floating point data values?
The sticking point seems to be the stipulation in the IEEE 754 standard that, when NaN is compared to any other value including itself, the answer is "unequal" (i.e. FALSE). This is clearly different from normal number comparisons and has repercussions for such things as indexing operations. Even so, this hardly applies to infinity values, which are single definite values. In fact, there is some encouraging talk in the Connect note on this issue that they might be supported ‘in the SQL Server 2008 timeframe’.
If didn’t happen; SQL 2008 doesn’t support NaN or infinity values, though one could be forgiven for thinking otherwise, based on the MSDN documentation for the FLOAT type, which states that "The behavior of float and real follows the IEEE 754 specification on approximate numeric data types". However, the truth is revealed in the XPath documentation, which states that "…float (53) is not exactly IEEE 754. For example, neither NaN (Not-a-Number) nor infinity is used…".
Is it really so hard to fix this problem the right way, and properly support in SQL Server the IEEE 754 standard for the floating point data type, NaNs, infinities and all? Oracle seems to have managed it quite nicely with its BINARY_FLOAT and BINARY_DOUBLE types, so it is technically possible.
We have an enterprise-class database that is marketed as being part of an ‘integrated’ Windows platform. Absurdly, we have .NET and XPath libraries that fully support the standard for floating point numbers, and we can’t even properly store these values, let alone query them, in the SQL Server database!