Construction and Practice of Qunar's Business Intelligence Platform

Qunar's BI platform has become a critical tool for data‑driven decision making, requiring easy‑to‑use drag‑and‑drop reporting, fast ad‑hoc queries, second‑level response times, and trustworthy metrics as the business scales.

1. Construction Phases

The platform evolved through three major stages since 2015: the original one‑stop report system (V1), a development stage with configurable reports and self‑service analysis (V2), and a system‑construction stage that introduced ad‑hoc queries, email reports, and a full‑stack data platform (V3).

2. Original Stage

In the early period, all data requests were handled by developers using a one‑stop pipeline: Hive parsed logs, performed ETL, loaded results into MySQL, and developers wrote backend and frontend code to render custom charts. This approach suffered from low efficiency, inconsistent code quality, and heavy duplication of similar reports.

3. Development Stage

To improve usability, the team rebuilt the reporting system (V2) with a configuration‑driven UI. Data developers exported ADS‑layer tables to PostgreSQL for analytical functions, while product users defined dimensions, metrics, and filters to generate reusable data units. Self‑service analysis was added, and an OLAP layer (including a custom OLAP engine) was introduced to support more complex queries.

3.1 Self‑service Analysis Architecture

The architecture required real‑time data ingestion, support for both offline and online tables, and sub‑second query performance. A combination of Impala+Kudu (for recent data) and Impala+HDFS (for historical data) was initially considered, but later benchmarks favored ClickHouse for its superior query speed on billion‑row datasets.

4. System‑Construction Stage

This stage focused on modularization and self‑service capabilities. Key components include:

Ad‑hoc query and email‑report modules that let users write SQL, with syntax checking, permission validation, and automatic result delivery.

Data‑report module built on a layered architecture: data source layer (MySQL, offline warehouse, metric system), ingestion layer (Waterdrop, Kafka), storage/engine layer (PostgreSQL for batch, ClickHouse for real‑time), data‑model layer (dimension/metric definitions), visualization layer (drag‑and‑drop charts, dashboards), and system‑management layer (permission integration, monitoring, lineage).

Drag‑and‑drop configuration stored as JSON, enabling front‑end rendering without developer intervention.

Advanced features such as multi‑metric calculations, alerting via QTalk/WeChat, and data lineage display for trustworthiness.

4.1 Engine Selection for Real‑time OLAP

After evaluating Druid, Impala, Doris, and ClickHouse, the team selected ClickHouse based on benchmark results showing sub‑second query latency on hundred‑billion‑row tables, efficient handling of joins, and low operational cost.

5. Impact and Future Directions

QBI now serves over a thousand daily active users across multiple business lines, supporting five major modules: ad‑hoc query, email reports, OLAP, data analysis, and data reporting. The platform delivers sub‑second response for most dashboards, enables self‑service data exploration, and plans to extend mobile BI, further abstraction of platform layers, and richer analysis scenarios such as retention and attribution.