Showing results for 
Search instead for 
Did you mean: 
Community Discussions
Connect with fellow community members to discuss general topics related to the Databricks platform, industry trends, and best practices. Share experiences, ask questions, and foster collaboration within the community.
Showing results for 
Search instead for 
Did you mean: 

log signature and input data for Spark LinearRegression


I am looking for a way to log my `` model with input and signature ata. The usual example that I found around are using sklearn and they can simply do 


# Log the model with signature and input example
signature = infer_signature(X_train, pd.DataFrame(y_train))
input_example = X_train.head(3)
mlflow.sklearn.log_model(rf_model, "rf_model", signature=signature, input_example=input_example)


but it doesn't work for `LinearRegression` because the "feature" column should be Spark `VectorUDT ` and it is not support by mlflow. This is how I generate my feature column


cat_input_cols = ["gender","occupation","zip_code","age_category"]
cat_index_output_cols = [x + '_index' for x in cat_input_cols]
ohe_output_cols = [x + '_ohe' for x in cat_input_cols]
stringIndexer = StringIndexer(inputCols=cat_input_cols, outputCols=cat_index_output_cols, handleInvalid="error", stringOrderType="alphabetDesc")
ohe_encoder = OneHotEncoder(inputCols=cat_index_output_cols, outputCols=ohe_output_cols)
 assembler = VectorAssembler(inputCols=features, outputCol="features")

pipeline = Pipeline(stages=[stringIndexer, ohe_encoder,assembler])
df_training_transformed =


now when I build my model I like to use `infer_signature` but I'm not able to find the right way to do it.


with mlflow.start_run(run_name="content_based_linReg") as run:
   # ---- set hyperparameters
    lr = LinearRegression(featuresCol=COL_FEATURES, labelCol=COL_LABEL)
    # ----  Split data into training and validation  sets
    (df_training_sampled, df_validation_sampled) = 
    #Create the model.
    lr_model =
    # Log the model parameters used for this run.
    mlflow.log_param("MAX_ITER", MAX_ITER)
    mlflow.log_param("REG_PARAM", REG_PARAM)
    mlflow.log_param("ELASTIC_NET_PARAM", ELASTIC_NET_PARAM)
    mlflow.log_param("FIT_INTERCEPT", FIT_INTERCEPT)

    # Run the model to create a prediction. Predict against the validation_df.
    df_validation_predictions = lr_model.transform(df_validation_sampled)

    # -->> this won't work
    # Log the model with signature and input example
    signature = infer_signature(df_training_sampled["features"], df_validation_predictions["prediction"])
    input_example ="features"),col("rating")).head(3)

    mlflow.spark.log_model(lr_model, "LinearRegression",sample_input=input_example,signature=signature)


here is the error I get when I call the `infer_signature`


Exception: Unsupported Spark Type '<class ''>', MLflow schema is only supported for scalar Spark types.


any idea how should I go about it? 


Community Manager
Community Manager

Hi @MohsenJThe error you’re encountering  infer_signature is due to the fact that it doesn’t directly support Spark’s VectorUDT type. However, we can work around this limitation.

Let’s create a custom signature for your LinearRegression model using the input features from your transformed DataFrame.

Here’s how you can proceed:

  1. Create a Custom Signature:

    • Instead of relying on automatic inference, we’ll manually construct a signature.
    • We’ll focus on the input features and ignore the VectorUDT column.
    • Assuming df_training_transformed is your transformed DataFrame, we’ll extract the relevant columns for the signature:
    from mlflow.models.signature import ModelSignature, Schema
    # Extract input features (excluding the VectorUDT column)
    input_columns = [col for col in df_training_transformed.columns if col != "features"]
    # Create a signature with input features
    signature = ModelSignature(inputs=Schema(input_columns))
  2. Log the Model with the Custom Signature:

    • After training your LinearRegression model, log it using MLflow.
    • Set the features column name to "features" (which is the default for LinearRegression).
    • Explicitly set the features column using setFeaturesCol.
    • Here’s an example:
    import mlflow
    from import LinearRegression
    # ... (your existing code)
    # Create and fit the Linear Regression model
    lr = LinearRegression(featuresCol="features", labelCol=COL_LABEL)
    lr_model =
    # Log the model with the custom signature
    mlflow.sklearn.log_model(lr_model, "linear_regression_model", signature=signature)
  3. Start an MLflow Run:

    • Use mlflow.start_run() to start an MLflow run.
    • Set hyperparameters and other relevant parameters.
    • Train your model and log the parameters.
  4. Log Model Parameters:

    • Use mlflow.log_param() to log any hyperparameters or other relevant parameters.

If you encounter any further issues or need additional assistance, feel free to ask! 😊



thanks @Kaniz_Fatma 

three clarification questions:

1.  wouldn't this cause issues when I load the model for inference? because in this case the model signature is different from the input?

2. I also need to log the signature of prediction output. should I just do

df_predictions = lr_model.transform(df_validation_sampled)
signature = ModelSignature(inputs=Schema(input_columns), output=Schema(df_predictions[["prediction]]))

3. to pass the input_sample, should I also just pass the rows of my dataset before transformation?


and one more question. you use mlflow.sklearn instead of mlflow.spark. why is that?

New Contributor III

Hey, at my team we are faced with the same problem.

Given that MLflow usage is part of the "Scalable Machine Learning with Apache Spark" course I got the impression that Spark ML models would work well with MLflow and Databricks model logging, but it's clearly not the case. I went back to the course material and, sure enough, when explaining the model registry it switches to using sklearn models. I gotta say, I feel a bit betrayed...

The solution given by @Kaniz_Fatma partially works but it is a hack and, as @MohsenJ said, it doesn't really cover the signature of the output. Is there a proper/more complete solution?

New Contributor II

Hi @Kaniz_Fatma@MohsenJ, did we get this working? I am facing a similar issue when trying to migrate one of the registered model to unity catalog. The infer signature from Mlflow doesn't seem to work as the input spark data frame contains features from VectorAssember that is recognised as type VectorUDT(). Model inputs contain unsupported Spark data types: [StructField('prd_dt', DateType(), False), StructField('features', VectorUDT(), True), StructField('features_scaled', VectorUDT(), True)]. Model signature is not logged. Failed to infer schema for Spark dataset. Exception: Unsupported Spark Type '<class ''>' for MLflow schema.

Have we found a proper solution for MLFlow to support datatypes other than scalar?

Would appreciate any inputs.

New Contributor III

@Abi105 I wasn't able to make it work, sorry


me neither. But the mlflow documentation suggests the new version of mlflow should be able to handle Array and Objects (dict). maybe that could help? I haven't tired it myself. 


Support for Array and Object types was introduced in MLflow version 2.10.0. These types will not be recognized in previous versions of MLflow. If you are saving a model that uses these signature types, you should ensure that any other environment that attempts to load these models has a version of MLflow installed that is at least 2.10.0.


Join 100K+ Data Experts: Register Now & Grow with Us!

Excited to expand your horizons with us? Click here to Register and begin your journey to success!

Already a member? Login and join your local regional user group! If there isn’t one near you, fill out this form and we’ll create one for you to join!