Scoring engines

DSS allows you to select various engines in order to perform scoring of your models. This allows for faster execution in some cases.

Note

Scoring engines are only used to actually predict rows. While they are strongly related to training engines, some models trained with one engine can be scored with another.

Engines for the scoring recipe

The following scoring engines are available:

• Local (DSS server only) scoring. This engine has two variants:

  • the Python engine provides wider compatibility but lower performance. Allows to compute Individual prediction explanations.

  • the Optimized scorer provides better performance and is automatically used whenever possible.

• Spark: the scoring is performed in a distributed fashion on a Spark cluster

• SQL (Regular): the model is converted to a SQL query and executed within a SQL database.

• SQL (Snowflake with Java UDF): the model uses Snowflake extended push-down. This provides much faster execution within Snowflake, and extended compatibility. Please see Snowflake for details

The selected engine can be adjusted in the scoring recipe editor. Only engines that are compatible with the selected model and input dataset will be available.

The default settings the following:

• If the model was trained with Spark MLLib, it will be scored with the Spark engine

• Else it will be scored with the Local engine. The optimized engine will be used if available.

If you do not wish to score your model with the “optimized” engine for some reason, you may select “Force original backend” in the scoring recipe editor to revert to the original backend.

Choosing SQL (regular) engine (if your scored dataset is stored in an SQL database and your model is compatible) will generate a request to score the dataset. Note that this may create very large requests for complex models.

Engines for the API node

To score rows using the API node, the “Local” engine is used. This engine has two variants:

• the Python engine provides wider compatibility but lower performance.

• the Optimized scorer provides better performance and is automatically used whenever possible.

The Optimized engine is enabled if you check the “Use Java” option in the endpoint settings.

In other words, only models for which one of “Local (Python)” or “Local (Optimized)” is available can be scored in the API node.

Compatibility matrix

The compatibility matrix for all DSS models is the following.

Local (Python) and Local (Optimized) engines are available both in scoring recipes and API node. Spark and SQL engines are only available for the scoring recipes.

Note

• For models trained with Python, the Optimized Local and Spark engines are only available if preprocessing is also compatible.

• The SQL engine is only available if preprocessing is also compatible.

Algorithms

Training engine	Algorithm	Local (Optimized)	Local (Python)	Spark	SQL (Snowflake with Java UDF)	SQL (Regular)
Python in-memory	Random forest	Yes	Yes	Yes	Yes	Yes (except for multiclass)
MLLib	Random forest	Yes	No	Yes	Yes	Yes (except for multiclass)
Python in-memory	Gradient Boosting	Yes	Yes	Yes	Yes	Yes (except for multiclass)
MLLib	Gradient Boosting (no multiclass)	Yes	No	Yes	Yes	Yes (except for multiclass)
Python in-memory	LightGBM	Yes	Yes	Yes	Yes	Yes (except for multiclass)
Python in-memory	XGBoost	Yes	Yes	Yes	Yes	Yes (except for multiclass)
Python in-memory	Extra Trees (Scikit)	Yes	Yes	Yes	Yes	Yes (except for multiclass)
Python in-memory	Decision Trees	Yes	Yes	Yes	Yes	Yes (no probas for multiclass)
MLLib	Decision Trees	Yes	No	Yes	Yes	Yes (no probas for multiclass)
Python in-memory	OLS, Lasso (non LARS), Ridge	Yes	Yes	Yes	Yes	Yes
Python in-memory	SGD	Yes	Yes	Yes	Yes	Yes
MLLib	Linear Regression	Yes	No	Yes	Yes	Yes
Python in-memory	Logistic Regression	Yes	Yes	Yes	Yes	Yes
MLLib	Logistic Regression	Yes	No	Yes	Yes	Yes
Python in-memory	Neural Networks	Yes	Yes	Yes	Yes	Yes
Python in-memory	Deep Neural Network	No	Yes	No	No	No
Python in-memory	Naive Bayes	No	Yes	No	No	No
MLLib	Naive Bayes	No	No	Yes	No	No
Python in-memory	K-nearest-neighbors	No	Yes	No	No	No
Python in-memory	SVM	No	Yes	No	No	No
Python in-memory	Custom models	No	Yes	No	No	No
MLLib	Custom models	No	No	Yes	No	No
	Ensemble model	No	Yes	No	No	No

Preprocessing

Local (Optimized)

The following preprocessing options are available for Local(Optimized)

• Numerical

  • No rescaling

  • Rescaling

  • Binning

  • Derivative features

  • Flag presence

  • Imputation

  • Drop row

  • Datetime cyclical encoding

• Categorical

  • Dummification

  • Target encoding (impact and GLMM)

  • Ordinal

  • Frequency

  • Flag presence

  • Hashing (MLLib only)

  • Impute

  • Drop row

• Text

  • Count vectorization

  • TF/IDF vectorization

  • Hashing (MLLib)

SQL (Snowflake with Java UDF)

The following preprocessing options are available for SQL (Snowflake with Java UDF) scoring :

• Numerical

  • No rescaling

  • Rescaling

  • Binning

  • Derivative features

  • Flag presence

  • Imputation

  • Drop row

  • Datetime cyclical encoding

• Categorical

  • Dummification

  • Target encoding (impact and GLMM)

  • Ordinal

  • Frequency

  • Flag presence

  • Hashing (MLLib only)

  • Impute

  • Drop row

• Text

  • Count vectorization

  • TF/IDF vectorization

  • Hashing (MLLib)

SQL (Regular)

The following preprocessing options are available for SQL (regular) scoring :

• Numerical

  • No rescaling

  • Rescaling

  • Binning

  • Flag presence

  • Imputation

  • Drop row

• Categorical

  • Dummification

  • Impact coding

  • Ordinal

  • Frequency

  • Flag presence

  • Impute

  • Drop row

Text is not supported

Limitations

SQL (regular) engine

The following limitations exist with SQL (regular) scoring:

• Some algorithms may not generate probabilities with SQL scoring (see table above)

• Conditional output columns are not generated with SQL scoring

• Preparation scripts are not compatible with SQL scoring

• Multiclass logistic regression and neural networks require the SQL dialect to support the GREATEST function.