perf(worker): 以有界微批次提升准入吞吐

原因:线上 P24 同构验收中,单任务同步复制提交与全局容量锁串行化,使两个 Worker 的 48 个执行槽只能维持约 10–14 个运行任务,最老等待超过 15 分钟。

影响:新增可配置的 1–32 条准入微批次,默认 8;租约、任务准入与唯一 River job 在一个有界事务内原子提交,并按确定顺序预锁任务和容量范围,避免整窗 48 条大事务和多 Dispatcher 死锁。并容忍 rebind 已被其他 Dispatcher 完成的幂等竞态。

验证:Go 全量测试、go vet、真实 PostgreSQL 跨 Store 集成测试、ShellCheck、迁移安全检查、OpenAPI、前端 lint/test/build、Compose/Kubernetes 渲染及人工发布脚本均通过。
This commit is contained in:
2026-08-01 20:26:47 +08:00
parent 92e328a575
commit c89c56ca65
13 changed files with 282 additions and 31 deletions
+5
View File
@@ -79,6 +79,7 @@ type Config struct {
AsyncWorkerHardLimit int
AsyncWorkerInstanceHardLimit int
AsyncWorkerRefreshIntervalSeconds int
AsyncAdmissionMicrobatchSize int
WorkerAutoscalingEnabled bool
WorkerReplicasNingbo int
WorkerReplicasHongkong int
@@ -186,6 +187,7 @@ func Load() Config {
),
AsyncWorkerInstanceHardLimit: envIntValidated("AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT", 32),
AsyncWorkerRefreshIntervalSeconds: envIntValidated("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS", 5),
AsyncAdmissionMicrobatchSize: envIntValidated("AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE", 8),
WorkerAutoscalingEnabled: env("AI_GATEWAY_WORKER_AUTOSCALING_ENABLED", "false") == "true",
WorkerReplicasNingbo: envOptionalIntValidated("AI_GATEWAY_WORKER_REPLICAS_NINGBO", 1),
WorkerReplicasHongkong: envOptionalIntValidated("AI_GATEWAY_WORKER_REPLICAS_HONGKONG", 1),
@@ -291,6 +293,9 @@ func (c Config) Validate() error {
if c.AsyncWorkerRefreshIntervalSeconds < 1 {
return errors.New("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS must be positive")
}
if c.AsyncAdmissionMicrobatchSize < 1 || c.AsyncAdmissionMicrobatchSize > 32 {
return errors.New("AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE must be between 1 and 32")
}
if c.WorkerReplicasNingbo < 0 || c.WorkerReplicasHongkong < 0 ||
c.WorkerMinReplicasNingbo < 0 || c.WorkerMinReplicasHongkong < 0 ||
c.WorkerMaxReplicasNingbo < c.WorkerMinReplicasNingbo ||