KGraph: Architecture, Performance, and Applications of Kuaishou's In‑House Graph Platform

Background KGraph is an in‑house graph platform developed by Kuaishou since late 2019, now used in social recommendation, e‑commerce recommendation, security and other scenarios.

Architecture KGraph adopts a directed heterogeneous property‑graph model, built on a self‑developed distributed key‑value storage layer (DBServer, Master, KNS) with PMem‑based persistent engine and a high‑performance RPC framework (KRPC). The storage layer provides read/write services across multiple regions and AZs.

Key challenges The system addresses the limitations of single‑machine memory and the poor performance of relational databases for large‑scale graph data, achieving per‑node throughput of tens of millions of QPS and sub‑millisecond latency.

Performance Benchmarks show single‑node read QPS reaching ten‑million level, overall system throughput up to 20 M QPS, and multi‑hop queries (2‑5 hops) completing within milliseconds, far outperforming traditional RDBMS solutions.

Applications KGraph powers real‑time social recommendation (friend suggestions, follow‑back), e‑commerce recommendation (product, live‑stream ads), and supports offline graph computation via Spark or Nebula.

Outlook Future work includes developing a native graph query engine, supporting additional storage back‑ends, and expanding to graph learning and knowledge‑graph use cases.