BCAI/easyai-ai-gateway
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

perf(worker): 以有界微批次提升准入吞吐

Browse Source 
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:
wangbo
2026-08-01 20:26:47 +08:00
parent 92e328a575
commit c89c56ca65
13 changed files with 282 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files
deploy/kubernetes/easyai-ai-gateway-cluster-release
+8 -1
View File
@@ -20,6 +20,7 @@ source "$config_file"
: "${AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT:=24}"
: "${AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT:=48}"
: "${AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT:=$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT}"
: "${AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE:=8}"
: "${AI_GATEWAY_WORKER_MEMORY_REQUEST_MIB:=1536}"
: "${AI_GATEWAY_WORKER_CPU_REQUEST_MILLICORES:=500}"
: "${AI_GATEWAY_DATABASE_MAX_CONNS:=32}"
@@ -71,6 +72,7 @@ for capacity_value in \
"$AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT" \
"$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT" \
"$AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT" \
"$AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE" \
"$AI_GATEWAY_WORKER_MEMORY_REQUEST_MIB" \
"$AI_GATEWAY_WORKER_CPU_REQUEST_MILLICORES" \
"$AI_GATEWAY_DATABASE_MAX_CONNS" \
@@ -113,6 +115,7 @@ done
AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT <= 256 &&
AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT <= 512 &&
AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT <= 512 &&
AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE <= 32 &&
AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT == AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT &&
AI_GATEWAY_WORKER_MEMORY_REQUEST_MIB >= 256 &&
AI_GATEWAY_WORKER_MEMORY_REQUEST_MIB <= 1900 &&
@@ -170,6 +173,7 @@ capacity_config_hash() {
"$AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT" \
"$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT" \
"$AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT" \
"$AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE" \
"$AI_GATEWAY_WORKER_MEMORY_REQUEST_MIB" \
"$AI_GATEWAY_WORKER_CPU_REQUEST_MILLICORES" \
"$AI_GATEWAY_DATABASE_MAX_CONNS" \
@@ -211,6 +215,7 @@ record_capacity_config() {
--arg autoscaling "$AI_GATEWAY_WORKER_AUTOSCALING_ENABLED" \
--arg instanceLimit "$AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT" \
--arg globalLimit "$AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT" \
--arg admissionMicrobatch "$AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE" \
--arg workerPool "$AI_GATEWAY_DATABASE_MAX_CONNS" \
--arg materialization "$AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY" \
--arg mediaRequest "$AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY" \
@@ -233,6 +238,7 @@ record_capacity_config() {
AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT:$instanceLimit,
AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT:$globalLimit,
AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT:$globalLimit,
AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE:$admissionMicrobatch,
AI_GATEWAY_WORKER_DATABASE_MAX_CONNS:$workerPool,
AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY:$materialization,
AI_GATEWAY_MEDIA_REQUEST_CONCURRENCY:$mediaRequest,
@@ -396,6 +402,7 @@ rollout_worker_site() {
"AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT=$AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT" \
"AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT=$AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT" \
"AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT=$AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT" \
"AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE=$AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE" \
"AI_GATEWAY_DATABASE_MAX_CONNS=$AI_GATEWAY_DATABASE_MAX_CONNS" \
"AI_GATEWAY_DATABASE_CRITICAL_MAX_CONNS=$AI_GATEWAY_DATABASE_CRITICAL_MAX_CONNS" \
"AI_GATEWAY_DATABASE_RIVER_MAX_CONNS=$AI_GATEWAY_DATABASE_RIVER_MAX_CONNS" \
@@ -617,7 +624,7 @@ apply_capacity_config() {
verify_site ningbo
wait_for_url 'public readiness' "$PUBLIC_BASE_URL/api/v1/readyz" '"ok":true'
record_capacity_config
echo "production_capacity=PASS config_hash=$(capacity_config_hash) ningbo_replicas=$AI_GATEWAY_WORKER_REPLICAS_NINGBO hongkong_replicas=$AI_GATEWAY_WORKER_REPLICAS_HONGKONG instance_limit=$AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT global_limit=$AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT worker_database_pool=$AI_GATEWAY_DATABASE_MAX_CONNS api_database_pool=$AI_GATEWAY_API_DATABASE_MAX_CONNS database_min_idle=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS database_max_idle_seconds=$AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS media_concurrency=$AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY worker_memory_request_mib=$AI_GATEWAY_WORKER_MEMORY_REQUEST_MIB worker_cpu_request_millicores=$AI_GATEWAY_WORKER_CPU_REQUEST_MILLICORES autoscaling=$AI_GATEWAY_WORKER_AUTOSCALING_ENABLED"
echo "production_capacity=PASS config_hash=$(capacity_config_hash) ningbo_replicas=$AI_GATEWAY_WORKER_REPLICAS_NINGBO hongkong_replicas=$AI_GATEWAY_WORKER_REPLICAS_HONGKONG instance_limit=$AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT global_limit=$AI_GATEWAY_ASYNC_WORKER_GLOBAL_HARD_LIMIT admission_microbatch=$AI_GATEWAY_ASYNC_ADMISSION_MICROBATCH_SIZE worker_database_pool=$AI_GATEWAY_DATABASE_MAX_CONNS api_database_pool=$AI_GATEWAY_API_DATABASE_MAX_CONNS database_min_idle=$AI_GATEWAY_DATABASE_MIN_IDLE_CONNS database_max_idle_seconds=$AI_GATEWAY_DATABASE_MAX_CONN_IDLE_SECONDS media_concurrency=$AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY worker_memory_request_mib=$AI_GATEWAY_WORKER_MEMORY_REQUEST_MIB worker_cpu_request_millicores=$AI_GATEWAY_WORKER_CPU_REQUEST_MILLICORES autoscaling=$AI_GATEWAY_WORKER_AUTOSCALING_ENABLED"
}
rollout_site() {