fix(openai-ws): ctx_pool 每账号连接上限系数默认值改为 5.0

backend/internal/config/config.go

@@ -1276,9 +1276,11 @@ type GatewayOpenAIWSConfig struct {
 	MaxIdlePerAccount  int `mapstructure:"max_idle_per_account"`
 	// DynamicMaxConnsByAccountConcurrencyEnabled: 是否按账号并发动态计算连接池上限
 	DynamicMaxConnsByAccountConcurrencyEnabled bool `mapstructure:"dynamic_max_conns_by_account_concurrency_enabled"`
-	// OAuthMaxConnsFactor: OAuth 账号连接池系数（effective=ceil(concurrency*factor)）
+	// OAuthMaxConnsFactor: OAuth 账号连接池系数（effective=ceil(concurrency*factor)，再受 max_conns_per_account 封顶）。
+	// ctx_pool 接入下每个客户端会话在整个生命周期（含轮次之间）持有一条上游连接，此上限限制的是同时持有连接的会话数，
+	// 在飞请求数另由账号并发槽限制；系数 1.0 会让存活会话数一到并发数就返回 1013 busy，默认 5.0。
 	OAuthMaxConnsFactor float64 `mapstructure:"oauth_max_conns_factor"`
-	// APIKeyMaxConnsFactor: API Key 账号连接池系数（effective=ceil(concurrency*factor)）
+	// APIKeyMaxConnsFactor: API Key 账号连接池系数，含义与 OAuthMaxConnsFactor 相同，默认 5.0。
 	APIKeyMaxConnsFactor  float64 `mapstructure:"apikey_max_conns_factor"`
 	DialTimeoutSeconds    int     `mapstructure:"dial_timeout_seconds"`
 	ReadTimeoutSeconds    int     `mapstructure:"read_timeout_seconds"`
@@ -2402,8 +2404,8 @@ func setDefaults() {
 	viper.SetDefault("gateway.openai_ws.min_idle_per_account", 4)
 	viper.SetDefault("gateway.openai_ws.max_idle_per_account", 12)
 	viper.SetDefault("gateway.openai_ws.dynamic_max_conns_by_account_concurrency_enabled", true)
-	viper.SetDefault("gateway.openai_ws.oauth_max_conns_factor", 1.0)
-	viper.SetDefault("gateway.openai_ws.apikey_max_conns_factor", 1.0)
+	viper.SetDefault("gateway.openai_ws.oauth_max_conns_factor", 5.0)
+	viper.SetDefault("gateway.openai_ws.apikey_max_conns_factor", 5.0)
 	viper.SetDefault("gateway.openai_ws.dial_timeout_seconds", 10)
 	viper.SetDefault("gateway.openai_ws.read_timeout_seconds", 900)
 	viper.SetDefault("gateway.openai_ws.write_timeout_seconds", 120)

backend/internal/config/config_test.go

@@ -452,11 +452,11 @@ func TestLoadDefaultOpenAIWSConfig(t *testing.T) {
 	if !cfg.Gateway.OpenAIWS.DynamicMaxConnsByAccountConcurrencyEnabled {
 		t.Fatalf("Gateway.OpenAIWS.DynamicMaxConnsByAccountConcurrencyEnabled = false, want true")
 	}
-	if cfg.Gateway.OpenAIWS.OAuthMaxConnsFactor != 1.0 {
-		t.Fatalf("Gateway.OpenAIWS.OAuthMaxConnsFactor = %v, want 1.0", cfg.Gateway.OpenAIWS.OAuthMaxConnsFactor)
+	if cfg.Gateway.OpenAIWS.OAuthMaxConnsFactor != 5.0 {
+		t.Fatalf("Gateway.OpenAIWS.OAuthMaxConnsFactor = %v, want 5.0", cfg.Gateway.OpenAIWS.OAuthMaxConnsFactor)
 	}
-	if cfg.Gateway.OpenAIWS.APIKeyMaxConnsFactor != 1.0 {
-		t.Fatalf("Gateway.OpenAIWS.APIKeyMaxConnsFactor = %v, want 1.0", cfg.Gateway.OpenAIWS.APIKeyMaxConnsFactor)
+	if cfg.Gateway.OpenAIWS.APIKeyMaxConnsFactor != 5.0 {
+		t.Fatalf("Gateway.OpenAIWS.APIKeyMaxConnsFactor = %v, want 5.0", cfg.Gateway.OpenAIWS.APIKeyMaxConnsFactor)
 	}
 	if cfg.Gateway.OpenAIWS.StickySessionTTLSeconds != 3600 {
 		t.Fatalf("Gateway.OpenAIWS.StickySessionTTLSeconds = %d, want 3600", cfg.Gateway.OpenAIWS.StickySessionTTLSeconds)

deploy/config.example.yaml

@@ -358,9 +358,13 @@ gateway:
     # 是否按账号并发动态计算连接池上限：
     # effective_max_conns = min(max_conns_per_account, ceil(account.concurrency * factor))
     dynamic_max_conns_by_account_concurrency_enabled: true
-    # 按账号类型分别设置系数（OAuth / API Key）
-    oauth_max_conns_factor: 1.0
-    apikey_max_conns_factor: 1.0
+    # 按账号类型分别设置系数（OAuth / API Key）。
+    # ctx_pool 接入下每个客户端会话在整个生命周期（轮次之间也不释放）持有一条上游连接，
+    # 该上限限制的是同时持有连接的会话数，在飞请求数另由账号并发限制；
+    # 系数 1.0 会让存活会话数一到并发数，新会话就收到 1013 upstream websocket is busy。
+    # 空闲连接数另受 max_idle_per_account 裁剪。
+    oauth_max_conns_factor: 5.0
+    apikey_max_conns_factor: 5.0
     dial_timeout_seconds: 10
     read_timeout_seconds: 900
     write_timeout_seconds: 120