Hologres Serverless Journey: How Alibaba Built Real-Time Data Warehousing

Hologres Serverless Journey

Hologres is Alibaba Cloud’s self‑developed, one‑stop real‑time data warehouse that integrates analytical services, supporting multi‑dimensional analysis, online services, lake‑warehouse integration, and vector computing.

Multi‑dimensional analysis – compatible with CK, Doris query scenarios.

Online services – high‑QPS KV and SQL point queries, non‑primary‑key point queries, row‑store with high availability.

Lake‑warehouse analysis – seconds‑level interactive queries on MaxCompute tables without data migration.

Vector computing – built‑in Proxima vector engine with performance surpassing open‑source vector databases.

Key Serverless Challenges and Technologies

Storage‑Compute Separation is a prerequisite for Serverless; data resides on a distributed file system while compute resources are elastic and can be launched on demand.

Shard Replica provides in‑memory replication of the latest data, enabling any compute node to access recent writes without flushing to disk, with maintenance cost only a fraction of the primary.

Isolation and Elasticity leverage the separated architecture to achieve strong resource isolation and automatic scaling, allowing seamless expansion from 500 QPS to 1000 QPS by simply adding resources.

High Availability ensures continuous connections during scaling or failover by routing queries to new compute instances without interruption.

Evolution Stages

Stage 1: Dedicated Instances – based on storage‑compute separation, users manually adjust instance specifications.

Stage 2: Shared Clusters – fully serverless lake‑warehouse clusters without storage, charging only for compute usage.

Stage 3: Master‑Slave Instances – one write‑able primary and multiple read‑only secondaries, providing strong isolation and rapid provisioning.

Stage 4: Compute‑Group Instances – multiple isolated compute groups within a single instance, unified routing via a gateway, enabling automatic scaling, isolation, and failover without application changes.

Customer Case: Logistics Real‑Time Warehouse

A logistics customer migrated from traditional warehouses to Hologres, moving through versions 1.0 (Kafka + Flink), 2.0 (master‑slave instances), and 3.0 (compute‑group instances), achieving up to 200 % performance improvement, 85 % latency reduction, and 100 % development efficiency gains.

Future Outlook

Alibaba Cloud aims for “Down To Zero” – no compute charges when idle and instant resource revival on demand – and to unify dedicated and shared instances into a single self‑service offering.