Brent wants us to blog about our favorite Connect items because Microsoft is currently listening to its customers and actually closing items. So it is the time to take advantage! The Connect item that I’m going to highlight was submitted by Niko Neugebauer (blog | twitter) and is – of course – about columnstore indexes:
One of the biggest advantages of columnstore indexes – aside from the huge compression ratio – is Batch Mode. With batch mode, certain operators in a query plan work on blocks of 1,000 rows at a time instead of doing operations row by row. This is a huge performance improvement: doing something once to 1,000 rows instead of doing it 1,000 times to an individual row tends to be faster. The downside is that it only works on tables with a columnstore index (clustered or non-clustered). If the table is in row storage, batch mode is not available and that’s what the connect item is about: make batch mode available in all scenarios. This can have some serious performance impact.
With SQL Server 2016, you can have batch mode on row storage table by using a little trick (detailed by Itzik Ben-Gan in SQLMag.com):
you can create a filtered non-clustered columnstore index on a single column. For example:
CREATE NONCLUSTERED COLUMNSTORE INDEX ix_dummy ON dbo.Table1(col1)
WHERE col1 = 0 AND col1 = 1;
Since col1 cannot be 0 and 1 at the same time, the index will have no data at all. So there is no performance overhead (no writes) and the optimizer will use a constant scan in the plan. However, since a columnstore index is involved, any query on Table1 can now run in batch mode. How crazy is that? This means SQL Server can already execute batch mode just fine on tables that are not really using a columnstore index, so there is no reason for not having this connect item fixed.
At the time of writing, the Connect items is over 2 years old but has only 36 votes. Time to change this! Please go to Connect and upvote it!