Columnstore data is physically stored in one or more segments (regular LOB allocation units) per column, and may also be partitioned in the usual way. Each segment contains roughly one million rows of highly-compressed values or value references (several compression techniques are available). A value reference links to an entry in one of up to two hash dictionaries.
Dictionaries are pinned in memory during query execution, with data value IDs from the segment being looked in the dictionary whenever execution requires the actual data value (this lookup is deferred for as long as possible for performance reasons).
Segments also have a header record containing metadata such as the minimum and maximum values stored in the segment. Information from the header can often be used to eliminate complete partitions from processing at execution time. Header record information is stored in the usual LOB data root structure, so eliminating a segment means the Storage Engine can skip reading the LOB data pages from physical storage entirely. Maximizing the potential for elimination can require careful design, including a dependency on the clustered index order at the time the Columnstore index is built.
Specific Plan Operators
SQL Server 2012 introduces a new execution mode called Batch Mode. In this mode, packets of roughly 1000 rows are passed between operators, significantly improving processor utilization efficiency. Within each packet, columnar data is represented as a vector. Not all plan operators support batch mode operation, but examples of those that do include Columnstore Index Scan, Hash Inner Join, Batch Hash Table Build, Bitmap Filter, Hash Aggregate (not scalar aggregates), Filter, and Compute Scalar (for projection and expression evaluation). Query execution plans have been enhanced to show estimated and actual execution mode.
There are a large number of restrictions in the first release, including constraints on allowable data types. Most common types are supported; unsupported data types include
DECIMAL with a precision greater than 18 digits,
UNIQUEIDENTIFIER, CLR types, and
Use of string types,
UNION ALL may result in significantly reduced performance (Row Mode execution), unless workarounds are employed that typically involve unusual syntax rewrites as shown in the linked articles in this section.
Remus Rusanu has blogged a great overview here.