feat: 添加5h窗口费用控制和会话数量限制

- 支持Anthropic OAuth/SetupToken账号的5h窗口费用阈值控制
- 支持账号级别的并发会话数量限制
- 使用Redis缓存窗口费用(30秒TTL)减少数据库压力
- 费用计算基于标准费用(不含账号倍率)
This commit is contained in:
shaw
2026-01-16 23:36:52 +08:00
parent 04811c00cb
commit 7379423325
21 changed files with 1329 additions and 53 deletions

View File

@@ -176,6 +176,7 @@ type GatewayService struct {
deferredService *DeferredService
concurrencyService *ConcurrencyService
claudeTokenProvider *ClaudeTokenProvider
sessionLimitCache SessionLimitCache // 会话数量限制缓存(仅 Anthropic OAuth/SetupToken
}
// NewGatewayService creates a new GatewayService
@@ -196,6 +197,7 @@ func NewGatewayService(
httpUpstream HTTPUpstream,
deferredService *DeferredService,
claudeTokenProvider *ClaudeTokenProvider,
sessionLimitCache SessionLimitCache,
) *GatewayService {
return &GatewayService{
accountRepo: accountRepo,
@@ -214,6 +216,7 @@ func NewGatewayService(
httpUpstream: httpUpstream,
deferredService: deferredService,
claudeTokenProvider: claudeTokenProvider,
sessionLimitCache: sessionLimitCache,
}
}
@@ -407,8 +410,12 @@ func (s *GatewayService) SelectAccountForModelWithExclusions(ctx context.Context
}
// SelectAccountWithLoadAwareness selects account with load-awareness and wait plan.
func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}) (*AccountSelectionResult, error) {
// metadataUserID: 原始 metadata.user_id 字段(用于提取会话 UUID 进行会话数量限制)
func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, groupID *int64, sessionHash string, requestedModel string, excludedIDs map[int64]struct{}, metadataUserID string) (*AccountSelectionResult, error) {
cfg := s.schedulingConfig()
// 提取会话 UUID用于会话数量限制
sessionUUID := extractSessionUUID(metadataUserID)
var stickyAccountID int64
if sessionHash != "" && s.cache != nil {
if accountID, err := s.cache.GetSessionAccountID(ctx, derefGroupID(groupID), sessionHash); err == nil {
@@ -527,7 +534,7 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
if len(routingAccountIDs) > 0 && s.concurrencyService != nil {
// 1. 过滤出路由列表中可调度的账号
var routingCandidates []*Account
var filteredExcluded, filteredMissing, filteredUnsched, filteredPlatform, filteredModelScope, filteredModelMapping int
var filteredExcluded, filteredMissing, filteredUnsched, filteredPlatform, filteredModelScope, filteredModelMapping, filteredWindowCost int
for _, routingAccountID := range routingAccountIDs {
if isExcluded(routingAccountID) {
filteredExcluded++
@@ -554,13 +561,18 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
filteredModelMapping++
continue
}
// 窗口费用检查(非粘性会话路径)
if !s.isAccountSchedulableForWindowCost(ctx, account, false) {
filteredWindowCost++
continue
}
routingCandidates = append(routingCandidates, account)
}
if s.debugModelRoutingEnabled() {
log.Printf("[ModelRoutingDebug] routed candidates: group_id=%v model=%s routed=%d candidates=%d filtered(excluded=%d missing=%d unsched=%d platform=%d model_scope=%d model_mapping=%d)",
log.Printf("[ModelRoutingDebug] routed candidates: group_id=%v model=%s routed=%d candidates=%d filtered(excluded=%d missing=%d unsched=%d platform=%d model_scope=%d model_mapping=%d window_cost=%d)",
derefGroupID(groupID), requestedModel, len(routingAccountIDs), len(routingCandidates),
filteredExcluded, filteredMissing, filteredUnsched, filteredPlatform, filteredModelScope, filteredModelMapping)
filteredExcluded, filteredMissing, filteredUnsched, filteredPlatform, filteredModelScope, filteredModelMapping, filteredWindowCost)
}
if len(routingCandidates) > 0 {
@@ -573,18 +585,25 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
if stickyAccount.IsSchedulable() &&
s.isAccountAllowedForPlatform(stickyAccount, platform, useMixed) &&
stickyAccount.IsSchedulableForModel(requestedModel) &&
(requestedModel == "" || s.isModelSupportedByAccount(stickyAccount, requestedModel)) {
(requestedModel == "" || s.isModelSupportedByAccount(stickyAccount, requestedModel)) &&
s.isAccountSchedulableForWindowCost(ctx, stickyAccount, true) { // 粘性会话窗口费用检查
result, err := s.tryAcquireAccountSlot(ctx, stickyAccountID, stickyAccount.Concurrency)
if err == nil && result.Acquired {
_ = s.cache.RefreshSessionTTL(ctx, derefGroupID(groupID), sessionHash, stickySessionTTL)
if s.debugModelRoutingEnabled() {
log.Printf("[ModelRoutingDebug] routed sticky hit: group_id=%v model=%s session=%s account=%d", derefGroupID(groupID), requestedModel, shortSessionHash(sessionHash), stickyAccountID)
// 会话数量限制检查
if !s.checkAndRegisterSession(ctx, stickyAccount, sessionUUID) {
result.ReleaseFunc() // 释放槽位
// 继续到负载感知选择
} else {
_ = s.cache.RefreshSessionTTL(ctx, derefGroupID(groupID), sessionHash, stickySessionTTL)
if s.debugModelRoutingEnabled() {
log.Printf("[ModelRoutingDebug] routed sticky hit: group_id=%v model=%s session=%s account=%d", derefGroupID(groupID), requestedModel, shortSessionHash(sessionHash), stickyAccountID)
}
return &AccountSelectionResult{
Account: stickyAccount,
Acquired: true,
ReleaseFunc: result.ReleaseFunc,
}, nil
}
return &AccountSelectionResult{
Account: stickyAccount,
Acquired: true,
ReleaseFunc: result.ReleaseFunc,
}, nil
}
waitingCount, _ := s.concurrencyService.GetAccountWaitingCount(ctx, stickyAccountID)
@@ -657,6 +676,11 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
for _, item := range routingAvailable {
result, err := s.tryAcquireAccountSlot(ctx, item.account.ID, item.account.Concurrency)
if err == nil && result.Acquired {
// 会话数量限制检查
if !s.checkAndRegisterSession(ctx, item.account, sessionUUID) {
result.ReleaseFunc() // 释放槽位,继续尝试下一个账号
continue
}
if sessionHash != "" && s.cache != nil {
_ = s.cache.SetSessionAccountID(ctx, derefGroupID(groupID), sessionHash, item.account.ID, stickySessionTTL)
}
@@ -699,15 +723,21 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
if ok && s.isAccountInGroup(account, groupID) &&
s.isAccountAllowedForPlatform(account, platform, useMixed) &&
account.IsSchedulableForModel(requestedModel) &&
(requestedModel == "" || s.isModelSupportedByAccount(account, requestedModel)) {
(requestedModel == "" || s.isModelSupportedByAccount(account, requestedModel)) &&
s.isAccountSchedulableForWindowCost(ctx, account, true) { // 粘性会话窗口费用检查
result, err := s.tryAcquireAccountSlot(ctx, accountID, account.Concurrency)
if err == nil && result.Acquired {
_ = s.cache.RefreshSessionTTL(ctx, derefGroupID(groupID), sessionHash, stickySessionTTL)
return &AccountSelectionResult{
Account: account,
Acquired: true,
ReleaseFunc: result.ReleaseFunc,
}, nil
// 会话数量限制检查
if !s.checkAndRegisterSession(ctx, account, sessionUUID) {
result.ReleaseFunc() // 释放槽位,继续到 Layer 2
} else {
_ = s.cache.RefreshSessionTTL(ctx, derefGroupID(groupID), sessionHash, stickySessionTTL)
return &AccountSelectionResult{
Account: account,
Acquired: true,
ReleaseFunc: result.ReleaseFunc,
}, nil
}
}
waitingCount, _ := s.concurrencyService.GetAccountWaitingCount(ctx, accountID)
@@ -748,6 +778,10 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
if requestedModel != "" && !s.isModelSupportedByAccount(acc, requestedModel) {
continue
}
// 窗口费用检查(非粘性会话路径)
if !s.isAccountSchedulableForWindowCost(ctx, acc, false) {
continue
}
candidates = append(candidates, acc)
}
@@ -765,7 +799,7 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
loadMap, err := s.concurrencyService.GetAccountsLoadBatch(ctx, accountLoads)
if err != nil {
if result, ok := s.tryAcquireByLegacyOrder(ctx, candidates, groupID, sessionHash, preferOAuth); ok {
if result, ok := s.tryAcquireByLegacyOrder(ctx, candidates, groupID, sessionHash, preferOAuth, sessionUUID); ok {
return result, nil
}
} else {
@@ -814,6 +848,11 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
for _, item := range available {
result, err := s.tryAcquireAccountSlot(ctx, item.account.ID, item.account.Concurrency)
if err == nil && result.Acquired {
// 会话数量限制检查
if !s.checkAndRegisterSession(ctx, item.account, sessionUUID) {
result.ReleaseFunc() // 释放槽位,继续尝试下一个账号
continue
}
if sessionHash != "" && s.cache != nil {
_ = s.cache.SetSessionAccountID(ctx, derefGroupID(groupID), sessionHash, item.account.ID, stickySessionTTL)
}
@@ -843,13 +882,18 @@ func (s *GatewayService) SelectAccountWithLoadAwareness(ctx context.Context, gro
return nil, errors.New("no available accounts")
}
func (s *GatewayService) tryAcquireByLegacyOrder(ctx context.Context, candidates []*Account, groupID *int64, sessionHash string, preferOAuth bool) (*AccountSelectionResult, bool) {
func (s *GatewayService) tryAcquireByLegacyOrder(ctx context.Context, candidates []*Account, groupID *int64, sessionHash string, preferOAuth bool, sessionUUID string) (*AccountSelectionResult, bool) {
ordered := append([]*Account(nil), candidates...)
sortAccountsByPriorityAndLastUsed(ordered, preferOAuth)
for _, acc := range ordered {
result, err := s.tryAcquireAccountSlot(ctx, acc.ID, acc.Concurrency)
if err == nil && result.Acquired {
// 会话数量限制检查
if !s.checkAndRegisterSession(ctx, acc, sessionUUID) {
result.ReleaseFunc() // 释放槽位,继续尝试下一个账号
continue
}
if sessionHash != "" && s.cache != nil {
_ = s.cache.SetSessionAccountID(ctx, derefGroupID(groupID), sessionHash, acc.ID, stickySessionTTL)
}
@@ -1081,6 +1125,107 @@ func (s *GatewayService) tryAcquireAccountSlot(ctx context.Context, accountID in
return s.concurrencyService.AcquireAccountSlot(ctx, accountID, maxConcurrency)
}
// isAccountSchedulableForWindowCost 检查账号是否可根据窗口费用进行调度
// 仅适用于 Anthropic OAuth/SetupToken 账号
// 返回 true 表示可调度false 表示不可调度
func (s *GatewayService) isAccountSchedulableForWindowCost(ctx context.Context, account *Account, isSticky bool) bool {
// 只检查 Anthropic OAuth/SetupToken 账号
if !account.IsAnthropicOAuthOrSetupToken() {
return true
}
limit := account.GetWindowCostLimit()
if limit <= 0 {
return true // 未启用窗口费用限制
}
// 尝试从缓存获取窗口费用
var currentCost float64
if s.sessionLimitCache != nil {
if cost, hit, err := s.sessionLimitCache.GetWindowCost(ctx, account.ID); err == nil && hit {
currentCost = cost
goto checkSchedulability
}
}
// 缓存未命中,从数据库查询
{
var startTime time.Time
if account.SessionWindowStart != nil {
startTime = *account.SessionWindowStart
} else {
startTime = time.Now().Add(-5 * time.Hour)
}
stats, err := s.usageLogRepo.GetAccountWindowStats(ctx, account.ID, startTime)
if err != nil {
// 失败开放:查询失败时允许调度
return true
}
// 使用标准费用(不含账号倍率)
currentCost = stats.StandardCost
// 设置缓存(忽略错误)
if s.sessionLimitCache != nil {
_ = s.sessionLimitCache.SetWindowCost(ctx, account.ID, currentCost)
}
}
checkSchedulability:
schedulability := account.CheckWindowCostSchedulability(currentCost)
switch schedulability {
case WindowCostSchedulable:
return true
case WindowCostStickyOnly:
return isSticky
case WindowCostNotSchedulable:
return false
}
return true
}
// checkAndRegisterSession 检查并注册会话,用于会话数量限制
// 仅适用于 Anthropic OAuth/SetupToken 账号
// 返回 true 表示允许在限制内或会话已存在false 表示拒绝(超出限制且是新会话)
func (s *GatewayService) checkAndRegisterSession(ctx context.Context, account *Account, sessionUUID string) bool {
// 只检查 Anthropic OAuth/SetupToken 账号
if !account.IsAnthropicOAuthOrSetupToken() {
return true
}
maxSessions := account.GetMaxSessions()
if maxSessions <= 0 || sessionUUID == "" {
return true // 未启用会话限制或无会话ID
}
if s.sessionLimitCache == nil {
return true // 缓存不可用时允许通过
}
idleTimeout := time.Duration(account.GetSessionIdleTimeoutMinutes()) * time.Minute
allowed, err := s.sessionLimitCache.RegisterSession(ctx, account.ID, sessionUUID, maxSessions, idleTimeout)
if err != nil {
// 失败开放:缓存错误时允许通过
return true
}
return allowed
}
// extractSessionUUID 从 metadata.user_id 中提取会话 UUID
// 格式: user_{64位hex}_account__session_{uuid}
func extractSessionUUID(metadataUserID string) string {
if metadataUserID == "" {
return ""
}
if match := sessionIDRegex.FindStringSubmatch(metadataUserID); len(match) > 1 {
return match[1]
}
return ""
}
func (s *GatewayService) getSchedulableAccount(ctx context.Context, accountID int64) (*Account, error) {
if s.schedulerSnapshot != nil {
return s.schedulerSnapshot.GetAccount(ctx, accountID)