Physical Data Warehouse Design - Data Warehouse Systems: Design and Implementation - page 266

hold in the ROLAP partition and the tables in it. The more relevant ones
are as follows:
￿ The partition cannot contain measures that use the MIN or MAX aggregate
functions.
￿ Each table in the schema of the ROLAP partition must be used only once.
￿ All table names in the partition schema must be qualified with the owner
name, for example, [dbo].[Customer] .
￿ All tables in the partition schema must have the same owner.
￿ The source columns of the partition measures must not be nullable.
MOLAP Storage
In the MOLAP storage mode, both the aggregations and a copy of the source
data are stored in a multidimensional structure. Such structures are highly
optimized to maximize query performance. Since a copy of the source data
resides in the multidimensional structure, queries can be processed without
accessing the source data of the partition.
Note however that data in a MOLAP partition reflect the most recently
processed state of a partition. Thus, when source data are updated, objects
in the MOLAP storage must be reprocessed to include the changes and
make them available to users. Changes can be processed from scratch or,
if possible, incrementally, as explained in Sect. 7.2 . This update can be
performed without taking the partition or cube off-line. However, if structural
changes to OLAP objects are performed, the cube must be taken off-line. In
these cases, it is recommended to update and process cubes on a staging
server.
HOLAP Storage
The HOLAP storage mode combines features of the previously explained
MOLAP and ROLAP modes. Like MOLAP, in HOLAP the aggregations of
the partition are stored in a multidimensional data structure. However, like
in ROLAP, HOLAP does not store a copy of the source data. Thus, if queries
only access summary data of a partition, HOLAP works like MOLAP very
eciently. Queries that need to access unaggregated source data must retrieve
it from the relational database and therefore will not be as fast as if it were
stored in a MOLAP structure. However, this can be solved if the query can
use cached data, that is, data that are stored in main memory rather than
on disk.
In summary, partitions stored as HOLAP are smaller than the equivalent
MOLAP partitions since they do not contain source data. On the other hand,
they can answer faster than ROLAP partitions for queries involving summary
data. Thus, this mode tries to capture the best of both worlds.