feat(gateway): 实现负载感知的账号调度优化

- 新增调度配置：粘性会话排队、兜底排队、负载计算、槽位清理
- 实现账号级等待队列和批量负载查询（Redis Lua 脚本）
- 三层选择策略：粘性会话优先 → 负载感知选择 → 兜底排队
- 后台定期清理过期槽位，防止资源泄漏
- 集成到所有网关处理器（Claude/Gemini/OpenAI）

backend/internal/repository/concurrency_cache.go

@@ -2,7 +2,9 @@ package repository
 
 import (
 	"context"
+	"errors"
 	"fmt"
+	"strconv"
 
 	"github.com/Wei-Shaw/sub2api/internal/service"
 	"github.com/redis/go-redis/v9"
@@ -27,6 +29,8 @@ const (
 	userSlotKeyPrefix = "concurrency:user:"
 	// 等待队列计数器格式: concurrency:wait:{userID}
 	waitQueueKeyPrefix = "concurrency:wait:"
+	// 账号级等待队列计数器格式: wait:account:{accountID}
+	accountWaitKeyPrefix = "wait:account:"
 
 	// 默认槽位过期时间（分钟），可通过配置覆盖
 	defaultSlotTTLMinutes = 15
@@ -112,33 +116,112 @@ var (
 			redis.call('EXPIRE', KEYS[1], ARGV[2])
 		end
 
-		return 1
-	`)
+			return 1
+		`)
+
+	// incrementAccountWaitScript - account-level wait queue count
+	incrementAccountWaitScript = redis.NewScript(`
+			local current = redis.call('GET', KEYS[1])
+			if current == false then
+				current = 0
+			else
+				current = tonumber(current)
+			end
+
+			if current >= tonumber(ARGV[1]) then
+				return 0
+			end
+
+			local newVal = redis.call('INCR', KEYS[1])
+
+			-- Only set TTL on first creation to avoid refreshing zombie data
+			if newVal == 1 then
+				redis.call('EXPIRE', KEYS[1], ARGV[2])
+			end
+
+			return 1
+		`)
 
 	// decrementWaitScript - same as before
 	decrementWaitScript = redis.NewScript(`
-		local current = redis.call('GET', KEYS[1])
-		if current ~= false and tonumber(current) > 0 then
-			redis.call('DECR', KEYS[1])
-		end
-		return 1
-	`)
+			local current = redis.call('GET', KEYS[1])
+			if current ~= false and tonumber(current) > 0 then
+				redis.call('DECR', KEYS[1])
+			end
+			return 1
+		`)
+
+	// getAccountsLoadBatchScript - batch load query (read-only)
+	// ARGV[1] = slot TTL (seconds, retained for compatibility)
+	// ARGV[2..n] = accountID1, maxConcurrency1, accountID2, maxConcurrency2, ...
+	getAccountsLoadBatchScript = redis.NewScript(`
+			local result = {}
+
+			local i = 2
+			while i <= #ARGV do
+				local accountID = ARGV[i]
+				local maxConcurrency = tonumber(ARGV[i + 1])
+
+				local slotKey = 'concurrency:account:' .. accountID
+				local currentConcurrency = redis.call('ZCARD', slotKey)
+
+				local waitKey = 'wait:account:' .. accountID
+				local waitingCount = redis.call('GET', waitKey)
+				if waitingCount == false then
+					waitingCount = 0
+				else
+					waitingCount = tonumber(waitingCount)
+				end
+
+				local loadRate = 0
+				if maxConcurrency > 0 then
+					loadRate = math.floor((currentConcurrency + waitingCount) * 100 / maxConcurrency)
+				end
+
+				table.insert(result, accountID)
+				table.insert(result, currentConcurrency)
+				table.insert(result, waitingCount)
+				table.insert(result, loadRate)
+
+				i = i + 2
+			end
+
+			return result
+		`)
+
+	// cleanupExpiredSlotsScript - remove expired slots
+	// KEYS[1] = concurrency:account:{accountID}
+	// ARGV[1] = TTL (seconds)
+	cleanupExpiredSlotsScript = redis.NewScript(`
+			local key = KEYS[1]
+			local ttl = tonumber(ARGV[1])
+			local timeResult = redis.call('TIME')
+			local now = tonumber(timeResult[1])
+			local expireBefore = now - ttl
+			return redis.call('ZREMRANGEBYSCORE', key, '-inf', expireBefore)
+		`)
 )
 
 type concurrencyCache struct {
-	rdb            *redis.Client
-	slotTTLSeconds int // 槽位过期时间（秒）
+	rdb                 *redis.Client
+	slotTTLSeconds      int // 槽位过期时间（秒）
+	waitQueueTTLSeconds int // 等待队列过期时间（秒）
 }
 
 // NewConcurrencyCache 创建并发控制缓存
 // slotTTLMinutes: 槽位过期时间（分钟），0 或负数使用默认值 15 分钟
-func NewConcurrencyCache(rdb *redis.Client, slotTTLMinutes int) service.ConcurrencyCache {
+// waitQueueTTLSeconds: 等待队列过期时间（秒），0 或负数使用 slot TTL
+func NewConcurrencyCache(rdb *redis.Client, slotTTLMinutes int, waitQueueTTLSeconds int) service.ConcurrencyCache {
 	if slotTTLMinutes <= 0 {
 		slotTTLMinutes = defaultSlotTTLMinutes
 	}
+	if waitQueueTTLSeconds <= 0 {
+		waitQueueTTLSeconds = slotTTLMinutes * 60
+	}
 	return &concurrencyCache{
-		rdb:            rdb,
-		slotTTLSeconds: slotTTLMinutes * 60,
+		rdb:                 rdb,
+		slotTTLSeconds:      slotTTLMinutes * 60,
+		waitQueueTTLSeconds: waitQueueTTLSeconds,
 	}
 }
 
@@ -155,6 +238,10 @@ func waitQueueKey(userID int64) string {
 	return fmt.Sprintf("%s%d", waitQueueKeyPrefix, userID)
 }
 
+func accountWaitKey(accountID int64) string {
+	return fmt.Sprintf("%s%d", accountWaitKeyPrefix, accountID)
+}
+
 // Account slot operations
 
 func (c *concurrencyCache) AcquireAccountSlot(ctx context.Context, accountID int64, maxConcurrency int, requestID string) (bool, error) {
@@ -225,3 +312,75 @@ func (c *concurrencyCache) DecrementWaitCount(ctx context.Context, userID int64)
 	_, err := decrementWaitScript.Run(ctx, c.rdb, []string{key}).Result()
 	return err
 }
+
+// Account wait queue operations
+
+func (c *concurrencyCache) IncrementAccountWaitCount(ctx context.Context, accountID int64, maxWait int) (bool, error) {
+	key := accountWaitKey(accountID)
+	result, err := incrementAccountWaitScript.Run(ctx, c.rdb, []string{key}, maxWait, c.waitQueueTTLSeconds).Int()
+	if err != nil {
+		return false, err
+	}
+	return result == 1, nil
+}
+
+func (c *concurrencyCache) DecrementAccountWaitCount(ctx context.Context, accountID int64) error {
+	key := accountWaitKey(accountID)
+	_, err := decrementWaitScript.Run(ctx, c.rdb, []string{key}).Result()
+	return err
+}
+
+func (c *concurrencyCache) GetAccountWaitingCount(ctx context.Context, accountID int64) (int, error) {
+	key := accountWaitKey(accountID)
+	val, err := c.rdb.Get(ctx, key).Int()
+	if err != nil && !errors.Is(err, redis.Nil) {
+		return 0, err
+	}
+	if errors.Is(err, redis.Nil) {
+		return 0, nil
+	}
+	return val, nil
+}
+
+func (c *concurrencyCache) GetAccountsLoadBatch(ctx context.Context, accounts []service.AccountWithConcurrency) (map[int64]*service.AccountLoadInfo, error) {
+	if len(accounts) == 0 {
+		return map[int64]*service.AccountLoadInfo{}, nil
+	}
+
+	args := []interface{}{c.slotTTLSeconds}
+	for _, acc := range accounts {
+		args = append(args, acc.ID, acc.MaxConcurrency)
+	}
+
+	result, err := getAccountsLoadBatchScript.Run(ctx, c.rdb, []string{}, args...).Slice()
+	if err != nil {
+		return nil, err
+	}
+
+	loadMap := make(map[int64]*service.AccountLoadInfo)
+	for i := 0; i < len(result); i += 4 {
+		if i+3 >= len(result) {
+			break
+		}
+
+		accountID, _ := strconv.ParseInt(fmt.Sprintf("%v", result[i]), 10, 64)
+		currentConcurrency, _ := strconv.Atoi(fmt.Sprintf("%v", result[i+1]))
+		waitingCount, _ := strconv.Atoi(fmt.Sprintf("%v", result[i+2]))
+		loadRate, _ := strconv.Atoi(fmt.Sprintf("%v", result[i+3]))
+
+		loadMap[accountID] = &service.AccountLoadInfo{
+			AccountID:          accountID,
+			CurrentConcurrency: currentConcurrency,
+			WaitingCount:       waitingCount,
+			LoadRate:           loadRate,
+		}
+	}
+
+	return loadMap, nil
+}
+
+func (c *concurrencyCache) CleanupExpiredAccountSlots(ctx context.Context, accountID int64) error {
+	key := accountSlotKey(accountID)
+	_, err := cleanupExpiredSlotsScript.Run(ctx, c.rdb, []string{key}, c.slotTTLSeconds).Result()
+	return err
+}

backend/internal/repository/concurrency_cache_benchmark_test.go

@@ -22,7 +22,7 @@ func BenchmarkAccountConcurrency(b *testing.B) {
 		_ = rdb.Close()
 	}()
 
-	cache, _ := NewConcurrencyCache(rdb, benchSlotTTLMinutes).(*concurrencyCache)
+	cache, _ := NewConcurrencyCache(rdb, benchSlotTTLMinutes, int(benchSlotTTL.Seconds())).(*concurrencyCache)
 	ctx := context.Background()
 
 	for _, size := range []int{10, 100, 1000} {

backend/internal/repository/concurrency_cache_integration_test.go

@@ -27,7 +27,7 @@ type ConcurrencyCacheSuite struct {
 
 func (s *ConcurrencyCacheSuite) SetupTest() {
 	s.IntegrationRedisSuite.SetupTest()
-	s.cache = NewConcurrencyCache(s.rdb, testSlotTTLMinutes)
+	s.cache = NewConcurrencyCache(s.rdb, testSlotTTLMinutes, int(testSlotTTL.Seconds()))
 }
 
 func (s *ConcurrencyCacheSuite) TestAccountSlot_AcquireAndRelease() {
@@ -218,6 +218,48 @@ func (s *ConcurrencyCacheSuite) TestWaitQueue_DecrementNoNegative() {
 	require.GreaterOrEqual(s.T(), val, 0, "expected non-negative wait count")
 }
 
+func (s *ConcurrencyCacheSuite) TestAccountWaitQueue_IncrementAndDecrement() {
+	accountID := int64(30)
+	waitKey := fmt.Sprintf("%s%d", accountWaitKeyPrefix, accountID)
+
+	ok, err := s.cache.IncrementAccountWaitCount(s.ctx, accountID, 2)
+	require.NoError(s.T(), err, "IncrementAccountWaitCount 1")
+	require.True(s.T(), ok)
+
+	ok, err = s.cache.IncrementAccountWaitCount(s.ctx, accountID, 2)
+	require.NoError(s.T(), err, "IncrementAccountWaitCount 2")
+	require.True(s.T(), ok)
+
+	ok, err = s.cache.IncrementAccountWaitCount(s.ctx, accountID, 2)
+	require.NoError(s.T(), err, "IncrementAccountWaitCount 3")
+	require.False(s.T(), ok, "expected account wait increment over max to fail")
+
+	ttl, err := s.rdb.TTL(s.ctx, waitKey).Result()
+	require.NoError(s.T(), err, "TTL account waitKey")
+	s.AssertTTLWithin(ttl, 1*time.Second, testSlotTTL)
+
+	require.NoError(s.T(), s.cache.DecrementAccountWaitCount(s.ctx, accountID), "DecrementAccountWaitCount")
+
+	val, err := s.rdb.Get(s.ctx, waitKey).Int()
+	if !errors.Is(err, redis.Nil) {
+		require.NoError(s.T(), err, "Get waitKey")
+	}
+	require.Equal(s.T(), 1, val, "expected account wait count 1")
+}
+
+func (s *ConcurrencyCacheSuite) TestAccountWaitQueue_DecrementNoNegative() {
+	accountID := int64(301)
+	waitKey := fmt.Sprintf("%s%d", accountWaitKeyPrefix, accountID)
+
+	require.NoError(s.T(), s.cache.DecrementAccountWaitCount(s.ctx, accountID), "DecrementAccountWaitCount on non-existent key")
+
+	val, err := s.rdb.Get(s.ctx, waitKey).Int()
+	if !errors.Is(err, redis.Nil) {
+		require.NoError(s.T(), err, "Get waitKey")
+	}
+	require.GreaterOrEqual(s.T(), val, 0, "expected non-negative account wait count after decrement on empty")
+}
+
 func (s *ConcurrencyCacheSuite) TestGetAccountConcurrency_Missing() {
 	// When no slots exist, GetAccountConcurrency should return 0
 	cur, err := s.cache.GetAccountConcurrency(s.ctx, 999)

backend/internal/repository/wire.go

@@ -10,7 +10,14 @@ import (
 // ProvideConcurrencyCache 创建并发控制缓存，从配置读取 TTL 参数
 // 性能优化：TTL 可配置，支持长时间运行的 LLM 请求场景
 func ProvideConcurrencyCache(rdb *redis.Client, cfg *config.Config) service.ConcurrencyCache {
-	return NewConcurrencyCache(rdb, cfg.Gateway.ConcurrencySlotTTLMinutes)
+	waitTTLSeconds := int(cfg.Gateway.Scheduling.StickySessionWaitTimeout.Seconds())
+	if cfg.Gateway.Scheduling.FallbackWaitTimeout > cfg.Gateway.Scheduling.StickySessionWaitTimeout {
+		waitTTLSeconds = int(cfg.Gateway.Scheduling.FallbackWaitTimeout.Seconds())
+	}
+	if waitTTLSeconds <= 0 {
+		waitTTLSeconds = cfg.Gateway.ConcurrencySlotTTLMinutes * 60
+	}
+	return NewConcurrencyCache(rdb, cfg.Gateway.ConcurrencySlotTTLMinutes, waitTTLSeconds)
 }
 
 // ProviderSet is the Wire provider set for all repositories