How Kuaishou Overcame the ‘Impossible Triangle’ of Performance, Flexibility, and Cost in Real‑Time Big Data Analytics

Project Background

Kuaishou’s content middle platform integrates content retrieval, diagnostics, and management, serving internal product, operation, and analysis teams with real‑time and hourly video and live‑stream data. The system consists of four core modules: retrieval, diagnostics, content pool, and foundational services.

Challenges

The platform faced a “impossible triangle” of high performance, high flexibility, and low cost while handling daily new content exceeding 40 million items and video play counts over 100 billion, requiring queries over half‑year data and 30+ consumption metrics.

Data volume: massive daily ingestion and long‑term storage.

Flexibility: need for arbitrary time, product, and page filters with dynamic ranking.

Cost: petabyte‑scale SSD storage and high‑throughput compute.

Technical Selection

After evaluating Elasticsearch and ClickHouse, ClickHouse was chosen as the primary analytical engine for its columnar storage, compression, and aggregation capabilities, especially for OLAP workloads.

Table Design – Solving Cost & Performance

To avoid the drawbacks of a single ultra‑wide table, a hybrid approach combining vertical partitioning (splitting columns into multiple tables) and vertical tables (storing key‑value pairs as rows) was adopted. This reduced storage bloat, improved query speed, and allowed independent maintenance of hot and cold data.

Query Performance Challenges

After table splitting, queries became more complex, involving dozens of tables and billions of rows. High‑cardinality aggregations (e.g., grouping by photo_id with billions of distinct values) and distributed joins caused memory pressure and network shuffling.

select sum(play_cnt) as order_cnt, photo_id
from photo_behav
where p_date between '20250101' and '20250103'
  and product = '极速版'
group by photo_id
order by order_cnt desc
limit 100;

Optimization Strategies

Data Pre‑Shuffle : During ETL, data is written with a distribution key so that rows sharing the same key land on the same ClickHouse node, enabling colocated queries without cross‑node shuffling.

Scatter‑Gather Execution Model : Queries are split (scatter) to worker nodes, executed locally, and results are gathered and merged, reducing network traffic.

Bitmap Aggregation : Using ClickHouse’s groupBitmap64State function to build Roaring Bitmaps for high‑cardinality sets, dramatically lowering memory usage.

Results

Performance improved from >20 seconds to <5 seconds for P80 queries, and required cluster size dropped from ~200 to ~70 machines. The hybrid table design and execution optimizations achieved a balanced solution to the performance‑flexibility‑cost triangle.

Conclusion

The case demonstrates that with careful data modeling, ClickHouse selection, and manual control over query execution, large‑scale real‑time analytics can meet stringent business requirements while keeping costs manageable.