Module | Core Feature | Detailed documentation |
Model Experiment | The model experiment panel tracks training parameters, metrics, and artifacts; it supports three types of no-code experiments: custom model experiments, agent experiments, and AutoML experiments (classification / regression / time series forecasting). | |
Feature Management | In Development Workbench, for each feature table, you can use the Feature Engineering SDK to perform operations such as creation, writing, reading, searching, synchronization, and consumption, and view and manage features on the Feature Management page to achieve unified feature management and consumption. | |
Model Management | You can register models in experiments by calling relevant MLflow functions, or perform visual model registration in experiment management. In the management interface, you can view key information of registered models and their associations with experiments/runs, services, and other entities. | |
Model Service | Deploy registered models as online inference APIs, supporting features such as version switching, traffic distribution, and resource monitoring. |
Module | Collaboration with data science |
Development Workbench | Provides a development environment where users can edit, debug, and run code in Development Workbench, and call MLflow and Feature Engineering APIs to perform operations such as feature table creation, modification, query, and deletion, model training, and model registration. |
Catalog | Unified storage and governance foundation for models and feature tables, providing permissions, tags, and lineage |
Workflow | Provides a workspace for automated workflows. After debugging code in Development Workbench, users can submit it to workflows to configure periodic scheduling, enabling automated and periodic model production. |
Data Quality | Model serving inference tables, feature tables, and training data tables can all initiate data quality jobs to view quality information such as field analysis, drift analysis, and model metrics. |
Type | Description |
Development Workbench | You must enable Development Workbench before you can use data science-related features. Development Workbench is the primary workspace for AI development, where users can edit, debug, and run code, and call MLflow and Feature Engineering APIs to perform operations such as feature table creation, modification, query, and deletion, model training, and model registration. |
Compute Resource Support | Compute resources of DataBuddy Development Workbench can be used for model training, experiment reporting, feature management, and model registration. For AutoML experiments, select the Machine Learning image when creating compute resources (not the default base image). |
MLflow Version | Experiment management is compatible with MLflow versions greater than or equal to 3.0.0 and less than 4.0.0. The relevant SDK is automatically installed by a pre-execution script when the Kernel starts, rather than being preinstalled in the image. After connecting to the Development Workbench runtime environment, run the following command to check: %pip list | grep mlflow Official documentation: MLflow documentation |
Offline Feature Store | Offline feature tables support only tables in Catalog. Tables with a primary key and a timestamp are automatically identified as feature tables. When registering a feature table, you must specify the primary key and timestamp key, which are then used for feature indexing. |
Online Feature Store | DataBuddy currently supports only PostgreSQL as the online feature store. To prepare, follow these steps: Go to the cloud database to create a PostgreSQL instance. The region and network must be consistent with the scheduling resource group and the engine in use. Add a PostgreSQL data source in **Platform Management > Data Source Management** of DataBuddy, and test connectivity to the scheduling resource group. After connecting to the remote kernel in the development workbench, verify whether PostgreSQL is connected: import psycopg2
PG_HOST = "myhost" # Example: "10.0.0.51"
PG_PORT = "myport" # Example: "5432"
PG_DB = "mydb"
PG_USER = "myuser"
PG_PWD = "mypwd" # If no password is set, comment out this line and the password parameter below.
try:
conn = psycopg2.connect(
host=PG_HOST,
port=PG_PORT,
dbname=PG_DB,
user=PG_USER,
password=PG_PWD,
connect_timeout=3,
)
conn.close()
print("PostgreSQL is reachable")
except Exception as e:
print("PostgreSQL connection failed:", e) |
Table Operation Permission | Grant permissions in **Compute Resources > Corresponding Resource > More > Permissions**, and grant permissions in DataBuddy Catalog > Corresponding Catalog / Schema / Table > Permissions. |
Model Operation Permission | Grant permissions to users in DataBuddy Catalog > Corresponding Model Catalog > Permissions. |
Feature Engineering Package | After connecting to the engine, run the following command to install the feature engineering package (version 0.2.3 or later required): %pip install "wedata3-feature-engineering[mlflow3]>=0.2.3" Run the following command to check the version: %pip list | grep wedata3-feature-engineering |
Environment Variable Settings | In DataBuddy, the MLflow / Feast service addresses, temporary credentials, and other settings are automatically injected by the pre-execution script when the Kernel starts, so users do not need to configure them manually. |
フィードバック