Skip to main content
POST
Submit an AutoML training job

Authorizations

Authorization
string
header
required

The access token received from the authorization server in the OAuth 2.0 flow.

Body

application/json

Request to train an AutoML model.

Required fields depend on problem_type. See each field's description for details. Invalid combinations return a 422 validation error.

name
string
required

Human-readable training run name.

domain
string
required

Business domain for governance.

owner
string
required

Username who owns this run.

source_table
string
required

Fully qualified Iceberg training table in catalog.schema.table format.

problem_type
enum<string>
required

Problem categories.

contextual_bandit is retired: kept only so persisted jobs still deserialize, rejected for new requests.

Available options:
binary_classification,
multiclass_classification,
regression,
ranking,
clustering,
anomaly_detection,
contextual_bandit
algorithm
enum<string>
required

AutoML algorithms.

The xgboost_*, sparkml_* and sklearn_* values are accepted for new training requests. The LightGBM, legacy isolation_forest and VW values are retired and kept only so persisted jobs still deserialize.

Available options:
lightgbm_classifier,
lightgbm_regressor,
lightgbm_ranker,
xgboost_classifier,
xgboost_regressor,
xgboost_ranker,
isolation_forest,
vw_classifier,
vw_regressor,
vw_contextual_bandit,
sparkml_logistic_regression,
sparkml_linear_regression,
sparkml_decision_tree_classifier,
sparkml_decision_tree_regressor,
sparkml_random_forest_classifier,
sparkml_random_forest_regressor,
sparkml_mlp_classifier,
sparkml_kmeans,
sklearn_isolation_forest
preset
enum<string>
default:balanced

Tuning presets shared across algorithms.

Available options:
fast,
balanced,
best_quality
label_column
string | null
positive_class_value

Value treated as the positive class for binary classification.

feature_columns
string[] | null

Override feature columns. If omitted, the job infers features from the table schema, excluding the label, time/group role columns, and any exclude_columns.

exclude_columns
string[] | null

Columns to drop from inferred features. For example, ID columns, entity-grain columns, or columns like trans_num. Ignored when the caller sets feature_columns.

group_column
string | null

Group/query identifier for ranking.

train_split_cutoff
string | null

Date cutoff string in International Organization for Standardization (ISO) 8601 format. Rows with the time column before the cutoff go to train, those at or after go to test. Defaults to an 80/20 random split when omitted.

time_column
string | null

Column used by train_split_cutoff. Required when you configure train_split_cutoff.

hyperparameter_overrides
Hyperparameter Overrides · object | null
resources
SparkResourceSpec · object | null

Optional Spark resource overrides for the training DAG.

Any field you omit auto-sizes from the dataset's row count, byte size, and feature width via automl_resource_sizing. A field you provide always wins over the derived value. Memory follows Spark's <n>g / <n>m format. For example, "8g".

enable_shap
boolean
default:true

Compute SHAP explanations. Applies to XGBoost classifiers and regressors; ignored otherwise.

regression_loss
enum<string> | null

Loss for xgboost_regressor. Defaults to squared.

Available options:
squared,
absolute,
poisson,
tweedie,
gamma,
quantile
quantile_alpha
number | null

Quantile to predict, between 0 and 1 (exclusive). Required when regression_loss is quantile.

Required range: 0 < x < 1
n_clusters
integer | null

Fixed number of clusters for sparkml_kmeans. When omitted, the job picks the best k between the preset's k_min and k_max by silhouette score.

Required range: 2 <= x <= 100
contamination
number | null

Expected share of anomalous rows for sklearn_isolation_forest, up to 0.5. Defaults to 0.01.

Required range: 0 < x <= 0.5
experiment_name
string | null
registered_model_name
string | null
feature_engineering_job_id
string<uuid> | null
trigger_now
boolean
default:true

Response

Job created with rendered DAG.

A persisted training job.

job_id
string<uuid>
required
name
string
required
domain
string
required
owner
string
required
source_table
string
required
problem_type
enum<string>
required

Problem categories.

contextual_bandit is retired: kept only so persisted jobs still deserialize, rejected for new requests.

Available options:
binary_classification,
multiclass_classification,
regression,
ranking,
clustering,
anomaly_detection,
contextual_bandit
algorithm
enum<string>
required

AutoML algorithms.

The xgboost_*, sparkml_* and sklearn_* values are accepted for new training requests. The LightGBM, legacy isolation_forest and VW values are retired and kept only so persisted jobs still deserialize.

Available options:
lightgbm_classifier,
lightgbm_regressor,
lightgbm_ranker,
xgboost_classifier,
xgboost_regressor,
xgboost_ranker,
isolation_forest,
vw_classifier,
vw_regressor,
vw_contextual_bandit,
sparkml_logistic_regression,
sparkml_linear_regression,
sparkml_decision_tree_classifier,
sparkml_decision_tree_regressor,
sparkml_random_forest_classifier,
sparkml_random_forest_regressor,
sparkml_mlp_classifier,
sparkml_kmeans,
sklearn_isolation_forest
preset
enum<string>
required

Tuning presets shared across algorithms.

Available options:
fast,
balanced,
best_quality
status
enum<string>
required

Lifecycle of an AutoML training job.

  • READY: DAG uploaded but not yet triggered. The user opted out of auto-trigger at create time. Manual trigger flips it to QUEUED.
  • QUEUED: DAG triggered, waiting for Airflow to pick it up.
  • RUNNING: Airflow has the run going.
  • COMPLETE/FAILED: terminal.
Available options:
ready,
queued,
running,
complete,
failed
created_at
string<date-time>
required
updated_at
string<date-time>
required
rendered_dag
string | null

The rendered DAG source. Populated on creation.

mlflow_run_id
string | null
mlflow_experiment_name
string | null
registered_model_name
string | null
airflow_dag_id
string | null
airflow_dag_url
string | null
error_message
string | null
feature_engineering_job_id
string<uuid> | null