Inside Alibaba AnalyticDB: Architecture, Core Technologies, and Real‑Time Data Warehouse Innovations

AnalyticDB Overview

AnalyticDB is Alibaba's self‑developed, PB‑scale real‑time data warehouse, supporting online analytics for e‑commerce, advertising, logistics, and entertainment. It has evolved from single‑node Oracle to a distributed, cloud‑native platform.

Challenges

Blurred line between online and offline processing; everything becomes a service.

High concurrency and low latency requirements.

Mixed workloads needing both data processing and interactive queries.

Fusion of structured and unstructured data analysis.

Overall Architecture

AnalyticDB is a multi‑tenant Cloud Native Realtime Data Warehouse. Each tenant DB consists of Front Node, Compute Node, and Buffer Node, with additional cluster‑level services.

DB‑Level Service Components

Front Node : Handles JDBC/ODBC access, authentication, SQL parsing, routing, optimizer, and MPP scheduling.

Compute Node : Executes MPP workers and stores data (row‑column hybrid, metadata, indexes).

Buffer Node : Manages real‑time writes and triggers index building.

Cluster‑Level Service Components

Management Console

Admin Service (billing, lifecycle, OpenAPI/InnerAPI)

Global Meta Service (metadata, partitioning, DDL)

Job Service (asynchronous jobs like index build, scaling)

Connector Service (data source integration)

Monitoring & Alerting Service

Resource Management Service (creation, scaling, DNS/SLB, health checks)

Data Model

Tables are organized into Fact Table Groups and Dimension Table Groups. Fact tables require hash partition columns and optional list sub‑partitions; dimension tables are smaller and stored in each Compute Node.

Massive Data Capabilities

20000+ tables in a single marketing DB.

~3 PB per DB, >100 million daily queries.

Clusters with >2000 nodes.

Write throughput up to 10 million TPS.

Complex queries with QPS >100.

Import & Export

Supports data import via Alibaba Cloud Data Transmission Service (DTS), DataWorks, MySQL JDBC, and a custom uploader for large files. Export to OSS and MaxCompute is done via INSERT‑FROM‑SELECT‑like operations.

Core Technologies

High‑Performance SQL Parser

AnalyticDB replaced Antlr/JavaCC with a self‑developed FastSQL parser, achieving 20‑30× speedups over traditional parsers on TPC‑DS benchmarks.

SELECT * FROM t1 t WHERE comm_week BETWEEN CAST(date_format(date_add('day', -day_of_week('20180605')), 'yyyyMMdd') AS BIGINT)

Xuanwu Storage Engine

Implements read/write separation, row‑column hybrid storage, and full‑column indexes (inverted, bitmap, interval tree). Real‑time data is written to Buffer Nodes, merged asynchronously, and indexed for millisecond‑level query response.

Read‑write separation ensures writes do not block reads, and replicas provide high availability.

Xihe Compute Engine

A large‑scale MPP DAG engine with both rule‑based and cost‑based optimizers, vectorized execution, and GPU acceleration. Supports mixed workloads, code generation, and dynamic statistics collection.

Optimizer

Column/partition pruning, predicate push‑down, join reordering.

Cost‑based plan selection with plan cache and literal classification.

Execution Engine

Pipeline DAG execution with stage/operator level preemption, CFS‑like scheduling for fast queries, and asynchronous index building.

GPU Acceleration

GPU Engine adds a Plan Rewriter, Task Manager, Code Generator (LLVM‑JIT to PTX), CUDA Manager, Data Manager, and VRAM Manager. It fuses operators, generates GPU kernels, and manages GPU memory using slab allocation, achieving sub‑second query latency for complex OLAP workloads.

Elastic Scaling

Online scaling (up/down, expand/shrink) is transparent to users. Resources can be added or removed without downtime, and data is automatically re‑balanced.

Summary

AnalyticDB’s key innovations include a high‑performance FastSQL parser, the Xuanwu storage engine with row‑column hybrid storage and full‑column indexes, the Xihe compute engine with MPP/DAG, CBO, vectorization, and GPU acceleration, and seamless elastic scaling for PB‑level real‑time analytics.