Known issues in Machine Learning Services

This article describes known problems or limitations with machine learning components that are provided as an option in SQL Server 2016 and SQL Server 2017.

The information here applies to all of the following, unless specifically indicated:

  • SQL Server 2016

    • R Services (In-Database)
    • Microsoft R Server (Standalone)
  • SQL Server 2017

    • Machine Learning Services for R (In-Database)
    • Machine Learning Services for Python (In-Database)
    • Machine Learning Server (Standalone)

Setup and configuration issues

For a description of processes and common questions that are related to initial setup and configuration, see Upgrade and installation FAQ. It contains information about upgrades, side-by-side installation, and installation of new R or Python components.

Unable to install Python components in offline installations of SQL Server 2017 CTP 2.0 or later

If you install a pre-release version of SQL Server 2017 on a computer without internet access, the installer might fail to display the page that prompts for the location of the downloaded Python components. In such an instance, you can install the Machine Learning Services feature, but not the Python components.

This issue is fixed in the release version. If you encounter this issue, as a workaround, you can temporarily enable internet access for the duration of the setup. This limitation does not apply to R.

Applies to: SQL Server 2017 with Python

Install the latest service release to ensure compatibility with Microsoft R Client

If you install the latest version of Microsoft R Client and use it to run R on SQL Server in a remote compute context, you might get an error like the following:

You are running version 9.x.x of Microsoft R Client on your computer, which is incompatible with Microsoft R Server version 8.x.x. Download and install a compatible version.

SQL Server 2016 requires that the R libraries on the client exactly match the R libraries on the server. The restriction has been removed for releases later than R Server 9.0.1. However, if you encounter this error, verify the version of the R libraries that's used by your client and the server and, if necessary, update the client to match the server version.

The version of R that is installed with SQL Server R Services is updated whenever a SQL Server service release is installed. To ensure that you always have the most up-to-date versions of R components, be sure to install all service packs.

To ensure compatibility with Microsoft R Client 9.0.0, install the updates that are described in this support article.

To avoid problems with R packages, you can also upgrade the version of the R libraries that are installed on the server, by changing to the Modern Lifecycle policy that's described in the next section. When you do so, the version of R that's installed with SQL Server is updated on the same schedule that updates are published for Microsoft R Server, which ensures that both server and client always have the latest releases of Microsoft R.

Applies to: SQL Server 2016 R Services, with R Server version 9.0.0 or earlier

Warning of incompatible version when you connect to an older version of SQL Server R Services from a client by using SQL Server 2017

When you run R code in a SQL Server 2016 compute context, and either of the following two statements is true, you might see an error like the following:

  • You installed R Server (Standalone) on a client computer by using the setup wizard for SQL Server 2017.
  • You installed Microsoft R Server by using the separate Windows installer.

You are running version 9.0.0 of Microsoft R Client on your computer, which is incompatible with the Microsoft R Server version 8.0.3. Download and install a compatible version.

You can use binding in Microsoft R Server 9.0 and later releases, to upgrade the R components in SQL Server 2016 instances. To determine if support for upgrades is available for your version of R Services, see Upgrade an instance of R Services using SqlBindR.exe.

Applies to: SQL Server 2016 R Services, with R Server version 9.0.0 or earlier

Setup for SQL Server 2016 service releases might fail to install newer versions of R components

When you install a cumulative update or install a service pack for SQL Server 2016 on a computer that is not connected to the internet, the setup wizard might fail to display the prompt that lets you update the R components by using downloaded CAB files. This failure typically occurs when multiple components were installed together with the database engine.

As a workaround, you can install the service release by using the command line and specifying the /MRCACHEDIRECTORY argument as shown in this example, which installs CU1 updates:

C:\<path to installation media>\SQLServer2016-KB3164674-x64.exe /Action=Patch /IACCEPTROPENLICENSETERMS /MRCACHEDIRECTORY=<path to CU1 CAB files>

To get the latest installers, see Install machine learning components without internet access.

Applies to: SQL Server 2016 R Services, with R Server version 9.0.0 or earlier

Launchpad services fails to start if the version is different from the R version

If you install SQL Server R Services separately from the database engine, and the build versions are different, you might see the following error in the System Event log:

The SQL Server Launchpad service failed to start due to the following error: The service did not respond to the start or control request in a timely fashion.

For example, this error might occur if you install the database engine by using the release version, apply a patch to upgrade the database engine, and then add the R Services feature by using the release version.

To avoid this problem, make sure that all components have the same version number. If you upgrade one component, be sure to apply the same upgrade to all other installed components.

To view a list of the R version numbers that are required for each release of SQL Server 2016, see Install R components without internet access.

Remote compute contexts are blocked by a firewall in SQL Server instances that are running on Azure virtual machines

If you have installed SQL Server 2017 on a Windows Azure virtual machine, you might not be able to use compute contexts that require the use of the virtual machine's workspace. The reason is that, by default, the firewall on Azure virtual machines includes a rule that blocks network access for local R user accounts.

As a workaround, on the Azure VM, open Windows Firewall with Advanced Security, select Outbound Rules, and disable the following rule: Block network access for R local user accounts in SQL Server instance MSSQLSERVER. You can also leave the rule enabled, but change the security property to Allow if secure.

Implied authentication in SQLEXPRESS

When you run R jobs from a remote data-science workstation by using Integrated Windows authentication, SQL Server uses implied authentication to generate any local ODBC calls that might be required by the script. However, this feature did not work in the RTM build of SQL Server Express Edition.

To fix the issue, we recommend that you upgrade to a later service release.

If you cannot upgrade, you can use a SQL login to run remote R jobs that might require embedded ODBC calls.

Applies to: SQL Server 2016 R Services Express Edition

Performance limits when R libraries are called from other R tools

It is possible to call the R tools and libraries that are installed for SQL Server from an external R application such as RGui. This call might be handy when you install new packages, or when you run ad hoc tests on very short code samples.

However, be aware that outside of SQL Server, performance might be limited. For example, even if you have purchased the Enterprise Edition of SQL Server, R runs in single-threaded mode when you run your R code by using external tools. Performance should be better if you run your R code by initiating a SQL Server connection and using sp_execute_external_script, which calls the R libraries for you.

  • Avoid calling the R libraries that are used by SQL Server from external R tools.
  • If you need to run extensive R code on the SQL Server computer without using SQL Server, install a separate instance of R, such as Microsoft R Client, and then ensure that your R development tools point to the new library.

For more information, see Create a Standalone R Server.

The R script is throttled due to resource governance default values

In Enterprise Edition, you can use resource pools to manage external script processes. In some early release builds, the maximum memory that could be allocated to the R processes was 20 percent. Therefore, if the server had 32 GB of RAM, the R executables (RTerm.exe and BxlServer.exe) could use a maximum of 6.4 GB in a single request.

If you encounter resource limitations, check the current default. If 20 percent is not enough, see the documentation for SQL Server on how to change this value.

Applies to: SQL Server 2016 R Services, Enterprise Edition

R code execution and package or function issues

This section contains known issues that are specific to running R on SQL Server, as well as some issues that are related to the R libraries and tools published by Microsoft, including RevoScaleR.

For additional known issues that might affect R solutions, go to the Microsoft R Server site.

Limitations on processor affinity for R jobs

In the initial release build of SQL Server 2016, you could set processor affinity only for CPUs in the first k-group. For example, if the server is a 2-socket machine with two k-groups, only processors from the first k-group are used for the R processes. The same limitation applies when you configure resource governance for R script jobs.

This issue is fixed in SQL Server 2016 Service Pack 1. We recommend that you upgrade to the latest service release.

Applies to: SQL Server 2016 R Services RTM version

Changes to column types cannot be performed when reading data in a SQL Server compute context

If your compute context is set to the SQL Server instance, you cannot use the colClasses argument (or other similar arguments) to change the data type of columns in your R code.

For example, the following statement would result in an error if the column CRSDepTimeStr is not already an integer:

data <- RxSqlServerData(sqlQuery = "SELECT CRSDepTimeStr, ArrDelay  FROM AirlineDemoSmall",
                                connectionString = connectionString,
                                colClasses = c(CRSDepTimeStr = "integer"))

As a workaround, you can rewrite the SQL query to use CAST or CONVERT and present the data to R by using the correct data type. In general, performance is better when you work with data by using SQL rather than by changing data in the R code.

Applies to: SQL Server 2016 R Services

Avoid clearing workspaces when you execute R code in a SQL Server compute context

If you use the R command to clear your workspace of objects while running R code in a SQL Server compute context, or if you clear the workspace as part of an R script called by using sp_execute_external_script, you might get this error:

workspace object 'revoScriptConnection' not found

revoScriptConnection is an object in the R workspace that contains information about an R session that is called from SQL Server. However, if your R code includes a command to clear the workspace (such as rm(list=ls())), all information about the session and other objects in the R workspace is cleared as well.

As a workaround, avoid indiscriminate clearing of variables and other objects while you're running R in SQL Server. Although clearing the workspace is common when working in the R console, it can have unintended consequences.

  • To delete specific variables, use the R remove function: remove('name1', 'name2', ...).
  • If there are multiple variables to delete, save the names of temporary variables to a list and perform periodic garbage collection.

Restrictions on data that can be provided as input to an R script

You cannot use in an R script the following types of query results:

  • Data from a Transact-SQL query that references AlwaysEncrypted columns.

  • Data from a Transact-SQL query that references masked columns.

    If you need to use masked data in an R script, a possible workaround is to make a copy of the data in a temporary table and use that data instead.

Arguments varsToKeep and varsToDrop are not supported for SQL Server data sources

When you use the rxDataStep function to write results to a table, using the varsToKeep and varsToDrop is a handy way of specifying the columns to include or exclude as part of the operation. However, these arguments are not supported for SQL Server data sources.

Limited support for SQL data types in sp_execute_external_script

Not all data types that are supported in SQL can be used in R. As a workaround, consider casting the unsupported data type to a supported data type before passing the data to sp_execute_external_script.

For more information, see R libraries and data types.

Possible string corruption

Any round trip of string data from Transact-SQL to R and then to Transact-SQL again can result in corruption. This is due to the different encodings used in R and in SQL Server, as well as the various collations and languages that are supported in R and Transact-SQL. Any string in a non-ASCII encoding can potentially be handled incorrectly.

When you send string data to R, convert it to an ASCII representation, if possible.

This limitation applies to data passed between SQL Server and Python as well. Multibyte characters should be passed as UTF-8 and stored as Unicode.

Only one value of type raw can be returned from sp_execute_external_script

When a binary data type (the R raw data type) is returned from R, the value must be sent in the output data frame.

With data types other than raw, you can return parameter values along with the results of the stored procedure simply by adding the OUTPUT keyword. For more information, see Return data by using OUTPUT parameters.

If you want to use multiple output sets that include values of type raw, one possible workaround is to do multiple calls of the stored procedure, or to send the result sets back to SQL Server by using ODBC.

Loss of precision

Because Transact-SQL and R support various data types, numeric data types can suffer loss of precision during conversion.

For more information about implicit data-type conversion, see R libraries and data types.

Variable scoping error when you use the transformFunc parameter: The sample data set for the analysis has no variables

To transform data while you are modeling, you can pass a transformFunc argument in a function such as rxLinmod or rxLogit. However, nested function calls can lead to scoping errors in the SQL Server compute context, even if the calls work correctly in the local compute context.

For example, assume that you have defined two functions, f and g, in your local global environment, and g calls f. In distributed or remote calls involving g, the call to g might fail because f cannot be found, even if you have passed both f and g to the remote call.

If you encounter this problem, you can work around the issue by embedding the definition of f inside your definition of g, anywhere before g would ordinarily call f.

For example:

f <- function(x) { 2*x * 3 }  
g <- function(y) {   
              a <- 10 * y  

To avoid the error, rewrite the definition as follows:

g <- function(y){  
              f <- function(x) { 2*x +3}  
              a <- 10 * y  

Data import and manipulation using RevoScaleR

When varchar columns are read from a database, white space is trimmed. To prevent this, enclose strings in non-white-space characters.

When functions such as rxDataStep are used to create database tables that have varchar columns, the column width is estimated based on a sample of the data. If the width can vary, it might be necessary to pad all strings to a common length.

Using a transform to change a variable's data type is not supported when repeated calls to rxImport or rxTextToXdf are used to import and append rows, combining multiple input files into a single .xdf file.

Limited support for rxExec

In SQL Server 2016, the rxExec function that's provided by the RevoScaleR package can be used only in single-threaded mode.

Parallelism for rxExec across multiple processes is planned for a future release.

Increase the maximum parameter size to support rxGetVarInfo

If you use data sets with extremely large numbers of variables (for example, over 40,000), set the max-ppsize flag when you start R to use functions such as rxGetVarInfo. The max-ppsize flag specifies the maximum size of the pointer protection stack.

If you are using the R console (for example, in rgui.exe or rterm.exe), you can set the value of max-ppsize to 500000 by typing:

R --max-ppsize=500000

If you are using the DevelopR environment, you can set the max-ppsize flag by making the following call to the RevoIDE executable:

RevoIDE.exe /RCommandLine --max-ppsize=500000  

Issues with the rxDTree function

The rxDTree function does not currently support in-formula transformations. In particular, using the F() syntax for creating factors on the fly is not supported. However, numeric data is automatically binned.

Ordered factors are treated the same as factors in all RevoScaleR analysis functions except rxDTree.

Revolution R Enterprise and Microsoft R Open

This section lists issues specific to R connectivity, development, and performance tools that are provided by Revolution Analytics. These tools were provided in earlier pre-release versions of SQL Server 2017.

In general, we recommend that you uninstall these previous versions and install the latest version of SQL Server or Microsoft R Server.

Running side-by-side versions of Revolution R Enterprise

Installing Revolution R Enterprise side by side with any version of R Services (In-Database) is not supported.

If you have a license to use a different version of Revolution R Enterprise, you must put it on a separate computer from both the SQL Server instance and any workstation that you want to use to connect to the SQL Server instance.

The use of an R productivity environment is not supported

Some pre-release versions of R Services (In-Database) included an R development environment for Windows that was created by Revolution Analytics. This tool is no longer provided, and it is not supported.

For compatibility with R Services (In-Database), we strongly recommend that you install Microsoft R Client or Microsoft R Server instead of the Revolution Analytics tools. R Tools for Visual Studio and Visual Studio Code also supports Microsoft R solutions.

Compatibility issues with SQLite ODBC driver and RevoScaleR

Revision 0.92 of the SQLite ODBC driver is incompatible with RevoScaleR. Revisions 0.88-0.91 and 0.93 and later are known to be compatible.

Python code execution or Python package issues

This section contains known issues that are specific to running Python on SQL Server, as well as issues that are related to the Python packages published by Microsoft, including revoscalepy and microsoftml.

See also

What's new in SQL Server 2016

Troubleshooting machine learning in SQL Server