Going in Production - Expert Cube Development with SSAS Multidimensional Models - page 309

In this example, we're running our partition management tasks in parallel with our
dimension processing, where we're running a Process Update on each dimension
that may have changed. After both of these operations have completed, we then use
another processing task to process, in parallel, all of the measure groups in our cube
with a Process Default so that new partitions are fully processed and any existing
partitions with invalid aggregations have their aggregations rebuilt.
Push-mode processing
So far, when we've processed dimensions or partitions, we've been sending XMLA
commands to Analysis Services to initiate this. Analysis Services then generates
whatever SQL is necessary and runs it against a relational data source. Finally, it
loads the data returned by these queries into the object that is being processed. This is
known as pull-mode processing: Analysis Services is pulling data from a data source.
There is also another, much less often used, method for processing objects: push-mode
processing. In push-mode processing, we can use Integration Services to push data
from a Data Flow into a partition or dimension using a Partition Processing or a
Dimension Processing destination. No SQL is generated by Analysis Services in this
case - we are pushing data into Analysis Services ourselves.
Push-mode processing is useful when we need to load data into Analysis Services
from data sources that are not officially supported and for which Analysis Services
would not be able to generate valid SQL (such as text files, or databases like Sybase
or MySQL). It is also an alternative to running a Process Add on a partition, or a
Process Add or a Process Update on a dimension, in situations where data needs to
be loaded into Analysis Services; however, like Process Add and Process Update,
regular use can lead to performance degradation, and so running a Process Full
occasionally is necessary.
Proactive caching
The proactive caching features in Analysis Services allow processing to be kicked
off automatically in response to changes in the relational data. When they were first
introduced in Analysis Services 2005, they were very heavily hyped, but the truth is
that they are almost never used in a production environment. This isn't because there
are any major problems with the functionality, more that there are easier ways of
handling the requirement for 'real-time' data.