perf: 精排 P99 延迟优化 (180ms → ≤50ms)

[JamesSunXX]

概述

精排 P99 延迟从峰刺时 180ms 优化至目标 ≤ 50ms。核心问题：Callback EAS 调用与主请求在服务端排队竞争 + 高 alloc rate 触发 GC assist 征用精排 goroutine。

变更内容

Phase 1: Callback 隔离 + 内存预分配

• Callback 限流: RateLimitQPS 配置 + limiter.Allow() 非阻塞限速
• 背压丢弃: channel 满时丢弃 callback（可配置 DropOnBackpressure）
• AlgoDataGenerator: 共享 userFeatures 引用，contextFeatures slice 预分配

Phase 2: 减少 Copy + Batch 优化

• EasyrecAlgoDataGenerator: 直接转移 slice 所有权替代 copy
• batchCount 默认 100 → 200（仍可配置覆盖）

Phase 3: Proto 优化 + Metrics

• proto.Buffer sync.Pool 复用（>1MB 不回池）
• callback_channel_pending prometheus gauge 监控背压
• limiter 支持配置热更新（SetLimit/SetBurst）

新增配置字段 (CallBackConfig)

{
  "WorkerPoolSize": 10,
  "RateLimitQPS": 200,
  "JitterMaxMs": 500,
  "DropOnBackpressure": true
}

测试

• go build ./... ✅
• go vet 通过（改动包）
• go test 通过（改动包）；EAS/sort/bloomfilter 失败为 pre-existing

灰度建议

1. Phase 1: 配置 RateLimitQPS=200, JitterMaxMs=500，观察 P99
2. Phase 2: 监控 EAS 单次调用 P99 在 batch=200 时 ≤ 30ms
3. Phase 3: 观察 alloc rate 下降至 ≤ 2GB/s

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

sch seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}