How EventHouse Redefines AI‑Native Event Data Platforms for the Agent Era

Background

In the AI‑agent era, data consumption is shifting from human‑centric SQL queries to programmatic agent access. Event data such as user behavior, transaction flows, system states, and IoT telemetry are often routed by EventBus or EventStreaming and then become “dark data” because downstream storage, governance, and deep analysis are handled by separate systems, resulting in T+1 latency and manual data stitching.

EventHouse Overview

EventHouse is Alibaba Cloud’s AI‑native data platform that acts as a unified integration bridge for agents. It blends the openness of a data lake with the reliability of a data warehouse, providing ACID transactions, schema management, and fine‑grained access control.

Layered Architecture

Integration Layer : Connects heterogeneous sources (Kafka, RocketMQ, MySQL, OSS, etc.) and normalizes them into a unified format.

Data Layer (EventStore) : Columnar, compressed storage optimized for JSON‑style event streams. Compared with traditional databases it saves >50% storage cost while exposing tables, views, and materialized views for SQL access.

Metadata Layer (Open Catalog) : Hive‑compatible catalog that automatically discovers schemas, tracks full‑lineage, and manages schema evolution. It can be federated with Spark, Flink, Presto and other engines via the Hive Metastore Thrift API.

Intelligent Query Engine : Provides a single SQL dialect for both batch and streaming queries, supports zero‑ETL cross‑source joins, pushes filters and computation down to external stores, and will add federated query capabilities.

Smart Analysis Layer

Luma Agent adds an AI semantic layer that maps natural‑language questions to the correct data fields using business‑level annotations stored in the catalog. The built‑in Luma DataAgent can monitor data, detect anomalies, plan analysis paths, and generate root‑cause reports autonomously. EventHouse also implements the Model Context Protocol (MCP), allowing any MCP‑compatible AI agent to invoke its query tools programmatically.

Key Capabilities

Unified SQL for streaming and batch workloads.

Zero‑ETL cross‑source analysis without data movement.

Computation push‑down for efficient query execution (e.g., filter push‑down to MySQL to avoid full‑table scans).

Future federated query support.

AI‑driven conversational and autonomous analytics via Luma Agent.

Native MCP support for programmatic access by AI agents.

Typical Use Cases

E‑commerce: Create a virtual view that joins real‑time payment events from RocketMQ with user profiles in MySQL, enabling analysts to query a single view without manual ETL.

IoT: Join device telemetry streams with RDS‑stored device inventories to build health dashboards, reducing latency from T+1 to near‑real‑time.

Risk‑control: An AI agent uses MCP to fetch historical and real‑time transaction data, performs multi‑dimensional correlation, and delivers instant risk assessments.

Technical Highlights

EventStore’s columnar compression reduces storage cost by more than 50% compared with traditional relational databases.

Open Catalog’s automatic schema discovery and lineage tracking shorten data‑find time from days to seconds and ensure safe schema evolution.

Zero‑ETL cross‑source joins allow SQL JOIN between internal tables and external data sources (e.g., OSS logs, external RDS dimension tables) without physical data movement.

Compute push‑down optimizes network traffic and query latency by executing filters and aggregations at the source.

Architecture Diagram