DeWu Reach Strategy Platform and HBase Buffer Pool Architecture

The Reach Strategy platform abstracts user engagement as tasks, strategies, and actions (push, coupon, message box, SMS, etc.). A task contains one or more strategies, each strategy defines a set of actions that are executed at the scheduled time.

The platform’s "buffer pool" (蓄水池) enhances algorithmic push capabilities by temporarily storing user data before batch processing, enabling the system to handle algorithmic personalization at the hundred‑million to billion scale.

Architecture diagrams illustrate the task‑subject structure (task → strategy → action) and the execution flow. The platform integrates AB testing, copy‑race, flow‑back analysis, and timing decisions to optimize user communication.

HBase is chosen as the storage layer for the buffer pool due to its massive, persistent storage, column‑family model, and horizontal scalability. It supports row‑key prefix scans, high concurrency, low latency, and sparse storage, which satisfy the platform’s requirement of processing billions of reads/writes per hour.

RowKey design combines a hash of (spuId+templateCode) with spuId, templateCode, and userId, ensuring even data distribution across regions and preventing hotspot issues. Pre‑splitting regions based on rowKey prefixes further balances load across HBase nodes.

The processing pipeline consists of producers (generating rowKeys), the HBase data pool (storing and partitioning data), and consumers (persistent threads that scan, parse, rate‑limit, and execute actions, then delete processed rows). The design allows dynamic horizontal scaling of both partitions and consumer threads.

Since the buffer pool went live, the platform can reliably handle algorithmic push volumes of up to a billion events within two hours, with plans to expand scenarios and user coverage.