OpenMLDB 0.4.0 Full-Process Features and Quick‑Start Guide for Building End‑to‑End Online AI Applications

OpenMLDB 0.4.0 brings extensive performance optimizations, including LLVM‑based JIT for various CPU architectures, and a multi‑level skip‑list storage engine that greatly improves read/write speed for time‑series data.

The release adds three major new capabilities: (1) unified online and offline storage with a single table view, (2) a high‑availability offline task manager that supports Spark‑based jobs via extended SQL commands, and (3) a complete end‑to‑end AI workflow that can be driven by SDK or CLI.

The article walks through three parts: new full‑process features, quick‑start procedures for the single‑node and cluster editions, and a hands‑on workshop that builds an end‑to‑end AI application using a Kaggle taxi‑duration dataset.

For the single‑node version, a pre‑compiled binary or Docker image can be started with a script that launches the NameServer, API Server, and a single Tablet. The cluster edition adds ZooKeeper for metadata management, multiple Tablets for replication, and the TaskManager service for fault‑tolerant job execution.

Users can create databases and tables via standard ANSI‑SQL, import offline data (CSV/Parquet) by switching to offline mode, perform feature extraction with window functions, export the resulting feature set, train models with LightGBM (or TensorFlow/PyTorch), and then deploy the feature‑SQL as an online service.

After deployment, online data can be “water‑filled” into the in‑memory store, and a lightweight HTTP/RPC server provides real‑time predictions using the trained model, achieving sub‑10‑ms latency for time‑series window aggregations.

The article concludes with a summary of the ten steps required to launch a full‑process AI application and invites readers to join the open‑source community on GitHub.