How Beike Unified Search & Recommendation Architecture Boosted Efficiency and Stability

Search & Recommendation Use Cases

Beike (the platform) connects three core entities—agents, houses, and customers—to support both search‑driven and recommendation‑driven scenarios. Typical interactions include keyword‑based searches (e.g., district or business circle) and passive homepage recommendations that surface relevant listings without an explicit query.

Search Platform Evolution

Stage 1 (2017): Simple Search Service

Implemented on SolrCloud with separate write and query services. The write service handled full and incremental data loads; the query service performed basic query parsing, recall, and ranking, exposing a unified API for read/write and configuration changes.

Stage 2 (Platform Phase)

Introduced MySQL binlog listeners for real‑time data sync and migrated the underlying engine to Elasticsearch. Query logic was split into a high‑level SearchService (parsing, recall, ranking) and a low‑level BasicSearch that directly queried ES, allowing lightweight services to bypass complex logic.

Stage 3 (Search Cloud Platform, 2019)

Automated the entire business‑onboarding workflow: self‑service data source registration, configuration, and one‑click deployment. Configuration data resides in MongoDB, enabling seamless propagation from development to QA and production. Visual tools for tokenization, data‑flow monitoring, SLA tracking, and automated testing reduced onboarding time from nine days to three, increased developer productivity six‑fold, and cut failure rates by 60%.

Stage 4 (Search Mid‑Platform, 2020‑present)

Unified gateway now handles authentication, rate limiting, circuit breaking, and routing. The query layer consists of six core modules: intent parsing, multi‑path recall (text, vector, business), fusion, coarse‑to‑fine ranking, re‑ranking, and recommendation reasoning. Vector (Milvus) and graph search capabilities were added alongside the traditional Elasticsearch engine.

Recommendation Platform Evolution

Content‑Based Recommendation

Early offline similarity calculations were stored in Redis, serving static “similar” or “popular” house listings.

Real‑Time Personalized Recommendation

Added user profiles, real‑time behavior logs, and feature engineering. Multi‑source recall (text, vector, business) is fused and filtered before serving.

Intelligent Platform Recommendation

Refactored into a layered architecture (application, compute, data, model). Introduced workflow‑based re‑ranking, allowing business rules to be composed via configuration rather than code, dramatically shortening development cycles.

Unified Search & Recommendation Architecture

A shared gateway and central control module orchestrate downstream services. The control module validates parameters, schedules strategies, caches intermediate results, and performs graceful degradation. If a downstream service times out, the control module falls back to default data sources (e.g., Elasticsearch or Redis) to maintain availability.

Control Service

Stateless and free of business logic; it manages request routing, timeout budgeting, and module‑level circuit breaking. It can cache NLU and reasoning results and dynamically adjust downstream timeouts based on remaining budget.

Recall Service

Constructs requests, applies NLU correction, enriches with user profiles, and performs multi‑path recall: text via Elasticsearch, vector via Milvus, business via custom APIs. Results are fused, filtered, and returned to the control module.

Re‑ranking Service

Executes business‑specific rules (e.g., top‑placement, deduplication) using a workflow engine that supports opt‑in/opt‑out rule composition, enabling most scenarios to be satisfied through configuration rather than code changes.

Key Operational Practices

Monitoring: Comprehensive metrics covering data‑flow latency, QPS, error rates, slow‑query thresholds, click‑through, exposure, and conversion rates, with alerts for abnormal spikes.

Data Validation: Automatic schema verification by pulling sample records before full ingestion to prevent mismatches.

Full‑Link Tracing: Elasticsearch‑based APM collects per‑module latency for end‑to‑end analysis.

Parallel Recall: Thread‑pool parallelism queries multiple engines (vector, text) concurrently and merges results.

Gateway: Powered by Zuul (Spring Cloud) for routing, authentication, rate limiting, and circuit breaking.

Stability Guarantees: Redundant deployments across two data centers, circuit breaking, rate limiting, and regular load‑testing ensure >99.99% availability.

Future Plans

Package generic strategy‑model combos as “strategy packages” for non‑core businesses to adopt quickly.

Build a one‑stop effectiveness optimization platform that integrates experiment, model management, feature management, and monitoring tools.

Architecture Diagrams