Design and Architecture of an Algorithm Business Platform for Rapid Online Service Development

The algorithm business platform is built to quickly create online algorithm services, accelerate business iteration, and enhance user experience; it comprises online services, data, algorithms, and data analysis, with this article focusing on service engineering that improves development efficiency through API‑based capability delivery and continuous API upgrades.

To meet diverse and rapidly changing algorithm requirements, the platform’s architecture decouples algorithm R&D from engineering R&D via clear APIs, reducing communication overhead and dramatically increasing iteration speed for scenario‑specific algorithm services.

Reliability and stability are foundational, achieved through thoughtful architectural design and a strict change‑release process that ensures high‑quality software delivery.

High availability is realized with micro‑service clusters, multi‑datacenter deployment, and distributed storage; the system supports multi‑level degradation, load‑based throttling, and overload protection to balance stability with algorithm performance during traffic spikes.

The platform’s main modules include an API layer, a recall engine (low‑latency massive user‑item and item‑item retrieval), a ranking engine (real‑time ML/DL model serving), a metadata middleware for asynchronous inter‑module communication, a feature‑reporting subsystem for massive real‑time data upload, and an AB‑testing platform.

The recall engine follows a two‑stage design: an initial portrait stage (user behavior, categories, material relations) and a feature stage (user, category, material, interaction dimensions), both executed in parallel asynchronous pipelines to reduce I/O blocking, leverage multi‑core resources, and boost recall efficiency.

Iterative optimizations revealed that excessive threading can degrade performance and that service splitting adds overhead; the final solution employs CompletableFuture‑based asynchronous parallelism, markedly improving recall throughput, resource utilization, and system latency.

The ranking engine supports multiple inference frameworks (Spark, TensorFlow) and models (GBDT, LR, Wide&Deep, DIN), balancing low latency with extensibility for online experiments.

The metadata middleware adopts asynchronous communication to avoid performance penalties from bulk synchronous data transfers, accepting added complexity for real‑time business requirements.

The feature‑reporting subsystem handles massive data streams under high QPS, tolerating slight delays but ensuring complete data delivery for model training through feature encoding, compression, and channel capacity optimization.

The platform extensively uses Java 8 streams and lambdas to reduce code volume and improve maintainability, aligning well with its logical requirements.

All services are stateless, facilitating horizontal scaling; state is persisted in storage solutions such as Jimdb, Elasticsearch, and Vearch, providing key‑value, vector, and tokenized retrieval capabilities, which demand careful schema and index design and deep understanding of each storage engine.

The AB‑test platform consists of three modules: configuration management, real‑time traffic splitting based on user/device attributes, and real‑time effect analysis (clicks, views, GMV) using offline and streaming analytics, all displayed on a unified dashboard to continuously improve algorithm performance.

These subsystems are tightly integrated through well‑defined interaction protocols, forming a cohesive whole rather than isolated components.

Looking ahead, the platform will evolve on the JDOS elastic cloud, leveraging source code and JSS object management for a full lifecycle of algorithm scenario development, testing, deployment, and effect analysis, thereby reducing operational complexity and enabling continuous integration and capability upgrades.