BehaviR: Mobile‑Side Real‑Time Stream Processing Framework

Background: With increasingly powerful mobile networks, users spend only seconds in an app, making it hard to provide valuable services. Xianyu built a flow‑control system and this article introduces the client‑side real‑time processing framework.

What the framework does: It handles user‑behavior streams, supplies high‑quality data, performs stateful real‑time computation, makes decisions and triggers user outreach. The overall architecture is illustrated in the original diagram.

Data supply: BehaviR’s data module collects page actions, network requests, storage operations, and persists them according to a JSON configuration, e.g.:

{
  "actionTypeIn":["leave"],
  "sceneIn":[
    "https://market.m.taobao.com/app/idleFish-F2e/idlefish-renting/home",
    "https://market.wapa.taobao.com/app/idleFish-F2e/idlefish-renting/home"
  ],
  "taskArray":[
    {
      "taskType":"py_backtrace",
      "pythonName":"test_cep_rent_rule2",
      "filter":[
        {"alias":"e1","actionType":"leave"},
        {"alias":"e2","actionType":"pv"}
      ]
    }
  ]
}

Computation: The client adopts a lightweight, Python‑based engine (Walle) to execute stateful stream calculations. Triggers are defined by "actionType" values; when a trigger matches, data is fed to the engine.

Implementation choice: Python was selected for rapid prototyping and because the internal Walle runtime is stable. Developers describe logic via a Python DSL, which is compiled into a computation graph.

Pattern('e1') \
.where(KVCondition('actionType','pv')) \
.and(KVCondition('scene','item_detail')) \
.followby('e2') \
.where(KVCondition('actionType','leave')) \
.and(KVCondition('scene','item_detail'))

The same logic can be expressed in the standard DSL:

EVENT: e1->e2
WHERE e1.extra_info.actionType = 'pv'
AND e1.extra_info.scene = 'item_detail'
AND e2.extra_info.actionType = 'leave'
AND e2.extra_info.scene = 'item_detail'

Output: After computation, results are routed to three channels – business rights delivery, model input, and downstream computation – using a unified protocol and a decision‑distribution module.

Effect and outlook: The first version runs on Xianyu with millisecond latency and zero server load. Future work includes handling process termination, out‑of‑order events, and more complex aggregations.