fix(worker): 拆分准入事务并收敛验收失败状态

将批量准入改为每任务独立事务,避免 P24 压测时 48 条任务共享长事务造成 transactionid 锁潮与 Worker 槽位空转。失败验收在保持 validation 的同时清理未提交任务,并记录精确压力门禁原因。统一生产 0+2 Worker 拓扑、运行时 ConfigMap 与发布配置,补充镜像预检诊断和回归测试。\n\n验证:Go 全量测试、go vet、真实 PostgreSQL 48 任务集成测试、迁移安全检查、发布脚本测试、bash -n、ShellCheck、Compose 配置均通过。
This commit is contained in:
2026-08-01 18:23:00 +08:00
parent 215b6b607f
commit edcbbd6a87
11 changed files with 194 additions and 100 deletions
@@ -6,8 +6,8 @@ DESIRED_STATE_DIR=/usr/local/share/easyai-ai-gateway-release/production
# Worker 容量只需修改本发布配置并执行获授权的 `capacity`,不再修改 Go 代码。
# 副本数属于 Deployment 配置,不能由容器内部环境变量改变;发布工具读取这两个变量并写入 replicas。
AI_GATEWAY_WORKER_REPLICAS_NINGBO=1
AI_GATEWAY_WORKER_REPLICAS_HONGKONG=1
AI_GATEWAY_WORKER_REPLICAS_NINGBO=0
AI_GATEWAY_WORKER_REPLICAS_HONGKONG=2
AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT=24
AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=48
AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT=48
@@ -23,10 +23,10 @@ AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS=30
AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY=24
AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY=24
AI_GATEWAY_WORKER_AUTOSCALING_ENABLED=false
AI_GATEWAY_WORKER_MIN_REPLICAS_NINGBO=1
AI_GATEWAY_WORKER_MIN_REPLICAS_NINGBO=0
AI_GATEWAY_WORKER_MIN_REPLICAS_HONGKONG=1
AI_GATEWAY_WORKER_MAX_REPLICAS_NINGBO=1
AI_GATEWAY_WORKER_MAX_REPLICAS_HONGKONG=1
AI_GATEWAY_WORKER_MAX_REPLICAS_NINGBO=0
AI_GATEWAY_WORKER_MAX_REPLICAS_HONGKONG=2
AI_GATEWAY_WORKER_TARGET_OUTSTANDING_PER_REPLICA=48
AI_GATEWAY_WORKER_SCALE_UP_WINDOW_SECONDS=20
AI_GATEWAY_WORKER_SCALE_DOWN_STABILIZATION_SECONDS=600