perf(queue): 按策略动态扩缩异步 Worker
将 River 执行容量按平台模型与用户组的有效并发策略动态调整,并为长任务续租、并发租约原子抢占和限流退避补充保护。\n\n统一平台模型限流继承语义,兼容历史 platformLimits/modelLimits,并为三个迁移图像模型建立独立并发租约。补充管理端显式继承/覆盖配置、指标、单元测试及隔离 PostgreSQL 验收。\n\n验证:go test ./... -count=1;pnpm lint;pnpm test;pnpm build;隔离 PostgreSQL 并发原子性/续租测试;128 任务与三分钟长任务动态 Worker 验收。
This commit is contained in:
@@ -29,6 +29,11 @@ IDENTITY_MODE=hybrid
|
||||
# - hold: reject new production generation before any upstream request; existing settlements continue.
|
||||
BILLING_ENGINE_MODE=observe
|
||||
|
||||
# River 执行 worker 会按平台模型和活跃用户组中更严格的 concurrent 总量每 5 秒重算。
|
||||
# hard limit 是单进程安全边界;平台模型/用户组 concurrency lease 才是业务并发真值。
|
||||
AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=2048
|
||||
AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS=5
|
||||
|
||||
# Unified identity business settings are managed in System Settings > Unified
|
||||
# Identity. Deployment only supplies the SecretStore and infrastructure timing.
|
||||
AI_GATEWAY_PUBLIC_BASE_URL=http://localhost:8088
|
||||
|
||||
Reference in New Issue
Block a user