fix(gateway): 修复粘性会话预取分组错配并优化并发等待热路径

This commit is contained in:
yangjianbo
2026-02-22 16:43:33 +08:00
parent a89477ddf5
commit 2ee6c26676
7 changed files with 121 additions and 32 deletions

View File

@@ -244,7 +244,12 @@ func (h *GatewayHandler) Messages(c *gin.Context) {
if sessionKey != "" {
sessionBoundAccountID, _ = h.gatewayService.GetCachedSessionAccountID(c.Request.Context(), apiKey.GroupID, sessionKey)
if sessionBoundAccountID > 0 {
prefetchedGroupID := int64(0)
if apiKey.GroupID != nil {
prefetchedGroupID = *apiKey.GroupID
}
ctx := context.WithValue(c.Request.Context(), ctxkey.PrefetchedStickyAccountID, sessionBoundAccountID)
ctx = context.WithValue(ctx, ctxkey.PrefetchedStickyGroupID, prefetchedGroupID)
c.Request = c.Request.WithContext(ctx)
}
}

View File

@@ -230,14 +230,31 @@ func (h *ConcurrencyHelper) AcquireAccountSlotWithWait(c *gin.Context, accountID
// waitForSlotWithPing waits for a concurrency slot, sending ping events for streaming requests.
// streamStarted pointer is updated when streaming begins (for proper error handling by caller).
func (h *ConcurrencyHelper) waitForSlotWithPing(c *gin.Context, slotType string, id int64, maxConcurrency int, isStream bool, streamStarted *bool) (func(), error) {
return h.waitForSlotWithPingTimeout(c, slotType, id, maxConcurrency, maxConcurrencyWait, isStream, streamStarted)
return h.waitForSlotWithPingTimeout(c, slotType, id, maxConcurrency, maxConcurrencyWait, isStream, streamStarted, false)
}
// waitForSlotWithPingTimeout waits for a concurrency slot with a custom timeout.
func (h *ConcurrencyHelper) waitForSlotWithPingTimeout(c *gin.Context, slotType string, id int64, maxConcurrency int, timeout time.Duration, isStream bool, streamStarted *bool) (func(), error) {
func (h *ConcurrencyHelper) waitForSlotWithPingTimeout(c *gin.Context, slotType string, id int64, maxConcurrency int, timeout time.Duration, isStream bool, streamStarted *bool, tryImmediate bool) (func(), error) {
ctx, cancel := context.WithTimeout(c.Request.Context(), timeout)
defer cancel()
acquireSlot := func() (*service.AcquireResult, error) {
if slotType == "user" {
return h.concurrencyService.AcquireUserSlot(ctx, id, maxConcurrency)
}
return h.concurrencyService.AcquireAccountSlot(ctx, id, maxConcurrency)
}
if tryImmediate {
result, err := acquireSlot()
if err != nil {
return nil, err
}
if result.Acquired {
return result.ReleaseFunc, nil
}
}
// Determine if ping is needed (streaming + ping format defined)
needPing := isStream && h.pingFormat != ""
@@ -286,15 +303,7 @@ func (h *ConcurrencyHelper) waitForSlotWithPingTimeout(c *gin.Context, slotType
case <-timer.C:
// Try to acquire slot
var result *service.AcquireResult
var err error
if slotType == "user" {
result, err = h.concurrencyService.AcquireUserSlot(ctx, id, maxConcurrency)
} else {
result, err = h.concurrencyService.AcquireAccountSlot(ctx, id, maxConcurrency)
}
result, err := acquireSlot()
if err != nil {
return nil, err
}
@@ -310,7 +319,7 @@ func (h *ConcurrencyHelper) waitForSlotWithPingTimeout(c *gin.Context, slotType
// AcquireAccountSlotWithWaitTimeout acquires an account slot with a custom timeout (keeps SSE ping).
func (h *ConcurrencyHelper) AcquireAccountSlotWithWaitTimeout(c *gin.Context, accountID int64, maxConcurrency int, timeout time.Duration, isStream bool, streamStarted *bool) (func(), error) {
return h.waitForSlotWithPingTimeout(c, "account", accountID, maxConcurrency, timeout, isStream, streamStarted)
return h.waitForSlotWithPingTimeout(c, "account", accountID, maxConcurrency, timeout, isStream, streamStarted, true)
}
// nextBackoff 计算下一次退避时间

View File

@@ -176,7 +176,7 @@ func TestWaitForSlotWithPingTimeout_AccountAndUserAcquire(t *testing.T) {
t.Run("account_slot_acquired_after_retry", func(t *testing.T) {
c, _ := newHelperTestContext(http.MethodPost, "/v1/messages")
streamStarted := false
release, err := helper.waitForSlotWithPingTimeout(c, "account", 101, 2, time.Second, false, &streamStarted)
release, err := helper.waitForSlotWithPingTimeout(c, "account", 101, 2, time.Second, false, &streamStarted, true)
require.NoError(t, err)
require.NotNil(t, release)
require.False(t, streamStarted)
@@ -188,7 +188,7 @@ func TestWaitForSlotWithPingTimeout_AccountAndUserAcquire(t *testing.T) {
t.Run("user_slot_acquired_after_retry", func(t *testing.T) {
c, _ := newHelperTestContext(http.MethodPost, "/v1/messages")
streamStarted := false
release, err := helper.waitForSlotWithPingTimeout(c, "user", 202, 3, time.Second, false, &streamStarted)
release, err := helper.waitForSlotWithPingTimeout(c, "user", 202, 3, time.Second, false, &streamStarted, true)
require.NoError(t, err)
require.NotNil(t, release)
release()
@@ -207,7 +207,7 @@ func TestWaitForSlotWithPingTimeout_TimeoutAndStreamPing(t *testing.T) {
helper := NewConcurrencyHelper(concurrency, SSEPingFormatNone, 5*time.Millisecond)
c, _ := newHelperTestContext(http.MethodPost, "/v1/messages")
streamStarted := false
release, err := helper.waitForSlotWithPingTimeout(c, "account", 101, 2, 130*time.Millisecond, false, &streamStarted)
release, err := helper.waitForSlotWithPingTimeout(c, "account", 101, 2, 130*time.Millisecond, false, &streamStarted, true)
require.Nil(t, release)
var cErr *ConcurrencyError
require.ErrorAs(t, err, &cErr)
@@ -218,7 +218,7 @@ func TestWaitForSlotWithPingTimeout_TimeoutAndStreamPing(t *testing.T) {
helper := NewConcurrencyHelper(concurrency, SSEPingFormatComment, 10*time.Millisecond)
c, rec := newHelperTestContext(http.MethodPost, "/v1/messages")
streamStarted := false
release, err := helper.waitForSlotWithPingTimeout(c, "account", 101, 2, 70*time.Millisecond, true, &streamStarted)
release, err := helper.waitForSlotWithPingTimeout(c, "account", 101, 2, 70*time.Millisecond, true, &streamStarted, true)
require.Nil(t, release)
var cErr *ConcurrencyError
require.ErrorAs(t, err, &cErr)
@@ -236,12 +236,29 @@ func TestWaitForSlotWithPingTimeout_AcquireError(t *testing.T) {
helper := NewConcurrencyHelper(concurrency, SSEPingFormatNone, 5*time.Millisecond)
c, _ := newHelperTestContext(http.MethodPost, "/v1/messages")
streamStarted := false
release, err := helper.waitForSlotWithPingTimeout(c, "account", 1, 1, 200*time.Millisecond, false, &streamStarted)
release, err := helper.waitForSlotWithPingTimeout(c, "account", 1, 1, 200*time.Millisecond, false, &streamStarted, true)
require.Nil(t, release)
require.Error(t, err)
require.Contains(t, err.Error(), "redis unavailable")
}
func TestAcquireAccountSlotWithWaitTimeout_ImmediateAttemptBeforeBackoff(t *testing.T) {
cache := &helperConcurrencyCacheStub{
accountSeq: []bool{false},
}
concurrency := service.NewConcurrencyService(cache)
helper := NewConcurrencyHelper(concurrency, SSEPingFormatNone, 5*time.Millisecond)
c, _ := newHelperTestContext(http.MethodPost, "/v1/messages")
streamStarted := false
release, err := helper.AcquireAccountSlotWithWaitTimeout(c, 301, 1, 30*time.Millisecond, false, &streamStarted)
require.Nil(t, release)
var cErr *ConcurrencyError
require.ErrorAs(t, err, &cErr)
require.True(t, cErr.IsTimeout)
require.GreaterOrEqual(t, cache.accountAcquireCalls, 1)
}
type helperConcurrencyCacheStubWithError struct {
helperConcurrencyCacheStub
err error

View File

@@ -264,7 +264,12 @@ func (h *GatewayHandler) GeminiV1BetaModels(c *gin.Context) {
if sessionKey != "" {
sessionBoundAccountID, _ = h.gatewayService.GetCachedSessionAccountID(c.Request.Context(), apiKey.GroupID, sessionKey)
if sessionBoundAccountID > 0 {
prefetchedGroupID := int64(0)
if apiKey.GroupID != nil {
prefetchedGroupID = *apiKey.GroupID
}
ctx := context.WithValue(c.Request.Context(), ctxkey.PrefetchedStickyAccountID, sessionBoundAccountID)
ctx = context.WithValue(ctx, ctxkey.PrefetchedStickyGroupID, prefetchedGroupID)
c.Request = c.Request.WithContext(ctx)
}
}