Deploy a Data Mining Solution to Previous Versions of SQL Server
Topic Status: Some information in this topic is preview and subject to change in future releases. Preview information describes new features or changes to existing features in Microsoft SQL Server 2016 Community Technology Preview 2 (CTP2).
This section describes known compatibility issues that may arise when you attempt to deploy a data mining model or data mining structure that was created in an instance of SQL Server 2016 Analysis Services (SSAS) to a database that uses SQL Server 2005 Analysis Services, or when you deploy models created in SQL Server 2005 to an instance of SQL Server 2016.
Deployment to an instance of SQL Server 2000 Analysis Services is not supported.
The Microsoft Time Series algorithm was enhanced in SQL Server 2008 by the addition of a second, complementary algorithm, ARIMA. For more information about the changes in the time series algorithm, see Microsoft Time Series Algorithm.
Therefore, time series mining models that use the new ARIMA algorithm may behave differently when deployed to an instance of SQL Server 2005 Analysis Services.
If you have explicitly set the parameter PREDICTION_SMOOTHING to control the mixture of ARTXP and ARIMA models in prediction, when you deploy this model to an instance of SQL Server 2005, Analysis Services will raise an error stating that the parameter is not valid. To prevent this error, you must delete the PREDICTION_SMOOTHING parameter and convert the models to a pure ARTXP model.
Conversely, if you deploy a time series model that was created using SQL Server 2005 Analysis Services to an instance of SQL Server 2016, when you open the mining model in SQL Server Data Tools (SSDT), the definition files are first converted to the new format, and two new parameters are added by default to all time series models. The parameter FORECAST_METHOD is added with the default value of MIXED, and the parameter PREDICTION_SMOOTHING is added with the default value of 0.5. However, the model will continue to use only ARTXP for forecasting until you reprocess the model. As soon as you reprocess the model, prediction changes to use both ARIMA and ARTXP.
Therefore, if you wish to avoid changing the model, you should only browse the model and never process it. Alternatively, you could explicitly set the FORECAST_METHOD or PREDICTION_SMOOTHING parameters.
For detailed information about configuring mixed models, see Microsoft Time Series Algorithm Technical Reference.
If the provider that is used for the model's data source is SQL Client Data Provider 10, you must also modify the data source definition to specify the previous version of the SQL Server Native Client. Otherwise, SQL Server Data Tools generates an error stating that the provider is not registered.
If you use SQL Server 2016 Analysis Services (SSAS) to create a mining structure that contains a holdout partition used for testing data mining models, the mining structure can be deployed to an instance of SQL Server 2005, but the partition information will be lost.
When you open the mining structure in SQL Server 2005 Analysis Services, SQL Server Data Tools raises an error, and then regenerates the structure to remove the holdout partition.
After the structure has been rebuilt, the size of the holdout partition is no longer available in the Properties window; however, the value <ddl100_100:HoldoutMaxPercent>30</ddl100_100:HoldoutMaxPercent>) may still be present in the ASSL script file.
If you use SQL Server 2016 Analysis Services (SSAS) to apply a filter to a mining model, the model can be deployed to an instance of SQL Server 2005, but the filter will not be applied.
When you open the mining model, SQL Server Data Tools (SSDT) raises an error, and then regenerates the model to remove the filter.
You cannot restore a database backup that was created in SQL Server 2016 to an instance of SQL Server 2005. If you do so, SQL Server Management Studio generates an error.
If you create a backup of a SQL Server 2005 Analysis Services database and restore this backup on an instance of SQL Server 2016, all time series models are modified as described in the previous section.