Issue with FeatureEngineeringClient().log_model()

[FedeRaimondi]

I am receiving a weird error when trying to log an xgboost model using feature engineering api.

I was able to log the model correctly with classic mlflow.xgboost.log_model() without any issues but when I switched to feature store recommended approach I am not able to correctly log it.

This is the error I get, quite weird since I am not passing any code_path argument:

TypeError: log_model() got an unexpected keyword argument 'code_path'

I am using:

- Runtime: 15.4.x-cpu-ml-scala2.12

- mlflow: '3.1.4'

- databricks-feature-store: '0.17.0'

- xgboost: '3.0.2'

Here below is the code (I believe I'm following all the right steps per documentation) :

mlflow.set_tracking_uri("databricks")
mlflow.set_registry_uri("databricks-uc")

# Read feature store from Unity Catalog
feature_store_name = "catalog.schema.name"
fe = FeatureEngineeringClient()
df_spark = fe.read_table(name=feature_store_name)

seed = 42
target = 'y'
features = ["list of features to use"]
columns_to_keep = ['CustomerID', target]
df_spark = df_spark.select(columns_to_keep)

# Create feature lookup
feature_lookups = [
    FeatureLookup(
      table_name=feature_store_name,
      feature_names=features,
      lookup_key=['CustomerID']
    )
  ]

# Create a training set
training_set = fe.create_training_set(df=df_spark, 
                                      feature_lookups=feature_lookups, 
                                      label=target,
                                      exclude_columns=['CustomerID'])

# Loading the training df
training_df = training_set.load_df()

# Split the data into training and test sets
training_df, test_df = training_df.randomSplit([0.8, 0.2], seed=seed)
training_df = training_df.toPandas()
test_df = test_df.toPandas()

with mlflow.start_run() as run:
    
    xgb_clf = XGBClassifier(colsample_bytree=0.05,
                            max_depth=3,
                            max_leaves=20,
                            eta=0.1,
                            n_estimators=10,
                            seed=seed)

    xgb_clf_model = xgb_clf.fit(training_df[features], training_df[target])

    # prediction
    train_pred = xgb_clf_model.predict_proba(training_df[features])[:, 1]
    test_pred = xgb_clf_model.predict_proba(test_df[features])[:, 1]

    # score
    auc_train = roc_auc_score(training_df[target], train_pred)
    auc_test = roc_auc_score(test_df[target], test_pred)

    # Log model
    fe.log_model(
        model=xgb_clf_model.get_booster(),
        artifact_path="model",
        flavor=mlflow.xgboost,
        training_set=training_set
        )

[FedeRaimondi · Accepted solution]

There is a typo in the libraries versions: I was using databricks-feature-engineering version 0.13, by downgrading to databricks-feature-engineering==0.12.1 (current stable version as of today: 4th August 2025) the code above functions as expected.

[szymon_dybczak]

Thanks for sharing @FedeRaimondi  solution with us.