From 74fcdf3d42b28a48446a6c5b1fda4e9a4ef23c49 Mon Sep 17 00:00:00 2001 From: cyh Date: Sat, 8 Aug 2026 00:29:22 +0800 Subject: [PATCH] fix(openai): preserve exponential capacity backoff --- backend/internal/handler/failover_loop.go | 22 +++++++++++++- .../internal/handler/failover_loop_test.go | 29 +++++++++++++++++++ .../handler/openai_chat_completions.go | 4 ++- .../handler/openai_gateway_handler.go | 8 +++-- 4 files changed, 59 insertions(+), 4 deletions(-) diff --git a/backend/internal/handler/failover_loop.go b/backend/internal/handler/failover_loop.go index 951a2f6b35..edbe4e232a 100644 --- a/backend/internal/handler/failover_loop.go +++ b/backend/internal/handler/failover_loop.go @@ -36,6 +36,9 @@ const ( maxSameAccountRetries = 3 // sameAccountRetryDelay 同账号重试间隔 sameAccountRetryDelay = 500 * time.Millisecond + // maxRequestScopedRetryDelay 限制请求级瞬时错误的指数退避上限，避免高重试配置 + // 将单次请求拖入分钟级等待。 + maxRequestScopedRetryDelay = 8 * time.Second // singleAccountBackoffDelay 单账号分组 503 退避重试固定延时。 // Service 层在 SingleAccountRetry 模式下已做充分原地重试（最多 3 次、总等待 30s）， // Handler 层只需短暂间隔后重新进入 Service 层即可。 @@ -52,6 +55,21 @@ const ( // 语义上等同于「无可用账号」：候选账号都不满足分组的利润约束。 const profitVetoExhaustedMessage = "No available accounts: all candidates rejected by group profit control" +func sameAccountRetryDelayFor(failoverErr *service.UpstreamFailoverError, retryCount int) time.Duration { + if failoverErr == nil || !failoverErr.RequestScopedTransient || retryCount <= 1 { + return sameAccountRetryDelay + } + + delay := sameAccountRetryDelay + for i := 1; i < retryCount; i++ { + if delay >= maxRequestScopedRetryDelay/2 { + return maxRequestScopedRetryDelay + } + delay *= 2 + } + return delay +} + // FailoverState 跨循环迭代共享的 failover 状态 type FailoverState struct { SwitchCount int @@ -149,13 +167,15 @@ func (s *FailoverState) HandleFailoverError( // 重试次数上限 retryLimit 由调用方传入（账号级 pool_mode_retry_count 配置）。 if failoverErr.RetryableOnSameAccount && s.SameAccountRetryCount[accountID] < retryLimit { s.SameAccountRetryCount[accountID]++ + retryDelay := sameAccountRetryDelayFor(failoverErr, s.SameAccountRetryCount[accountID]) logger.FromContext(ctx).Warn("gateway.failover_same_account_retry", zap.Int64("account_id", accountID), zap.Int("upstream_status", failoverErr.StatusCode), zap.Int("same_account_retry_count", s.SameAccountRetryCount[accountID]), zap.Int("same_account_retry_max", retryLimit), + zap.Duration("retry_delay", retryDelay), ) - if !sleepWithContext(ctx, sameAccountRetryDelay) { + if !sleepWithContext(ctx, retryDelay) { return FailoverCanceled } return FailoverContinue diff --git a/backend/internal/handler/failover_loop_test.go b/backend/internal/handler/failover_loop_test.go index 34644711eb..e65b43c172 100644 --- a/backend/internal/handler/failover_loop_test.go +++ b/backend/internal/handler/failover_loop_test.go @@ -31,6 +31,35 @@ func (m *mockTempUnscheduler) TempUnscheduleRetryableError(_ context.Context, ac m.calls = append(m.calls, tempUnscheduleCall{accountID: accountID, failoverErr: failoverErr}) } +func TestSameAccountRetryDelayFor(t *testing.T) { + capacityErr := &service.UpstreamFailoverError{RequestScopedTransient: true} + + for _, tc := range []struct { + name string + retryCount int + want time.Duration + }{ + {name: "first retry", retryCount: 1, want: 500 * time.Millisecond}, + {name: "second retry", retryCount: 2, want: time.Second}, + {name: "third retry", retryCount: 3, want: 2 * time.Second}, + {name: "fourth retry", retryCount: 4, want: 4 * time.Second}, + {name: "fifth retry", retryCount: 5, want: 8 * time.Second}, + {name: "capped retry", retryCount: 10, want: 8 * time.Second}, + } { + t.Run(tc.name, func(t *testing.T) { + require.Equal(t, tc.want, sameAccountRetryDelayFor(capacityErr, tc.retryCount)) + }) + } + + t.Run("non request scoped errors keep fixed delay", func(t *testing.T) { + require.Equal(t, 500*time.Millisecond, sameAccountRetryDelayFor(&service.UpstreamFailoverError{}, 10)) + }) + + t.Run("nil error keeps fixed delay", func(t *testing.T) { + require.Equal(t, 500*time.Millisecond, sameAccountRetryDelayFor(nil, 10)) + }) +} + // --------------------------------------------------------------------------- // Helper // --------------------------------------------------------------------------- diff --git a/backend/internal/handler/openai_chat_completions.go b/backend/internal/handler/openai_chat_completions.go index 9715d7917e..42dc13112a 100644 --- a/backend/internal/handler/openai_chat_completions.go +++ b/backend/internal/handler/openai_chat_completions.go @@ -288,16 +288,18 @@ func (h *OpenAIGatewayHandler) ChatCompletions(c *gin.Context) { retryLimit := account.GetPoolModeRetryCount() if sameAccountRetryCount[account.ID] < retryLimit { sameAccountRetryCount[account.ID]++ + retryDelay := sameAccountRetryDelayFor(failoverErr, sameAccountRetryCount[account.ID]) reqLog.Warn("openai_chat_completions.pool_mode_same_account_retry", zap.Int64("account_id", account.ID), zap.Int("upstream_status", failoverErr.StatusCode), zap.Int("retry_limit", retryLimit), zap.Int("retry_count", sameAccountRetryCount[account.ID]), + zap.Duration("retry_delay", retryDelay), ) select { case <-c.Request.Context().Done(): return - case <-time.After(sameAccountRetryDelay): + case <-time.After(retryDelay): } continue } diff --git a/backend/internal/handler/openai_gateway_handler.go b/backend/internal/handler/openai_gateway_handler.go index b787d0f976..77fd1e8ea7 100644 --- a/backend/internal/handler/openai_gateway_handler.go +++ b/backend/internal/handler/openai_gateway_handler.go @@ -610,16 +610,18 @@ func (h *OpenAIGatewayHandler) Responses(c *gin.Context) { retryLimit := account.GetPoolModeRetryCount() if sameAccountRetryCount[account.ID] < retryLimit { sameAccountRetryCount[account.ID]++ + retryDelay := sameAccountRetryDelayFor(failoverErr, sameAccountRetryCount[account.ID]) reqLog.Warn("openai.pool_mode_same_account_retry", zap.Int64("account_id", account.ID), zap.Int("upstream_status", failoverErr.StatusCode), zap.Int("retry_limit", retryLimit), zap.Int("retry_count", sameAccountRetryCount[account.ID]), + zap.Duration("retry_delay", retryDelay), ) select { case <-c.Request.Context().Done(): return - case <-time.After(sameAccountRetryDelay): + case <-time.After(retryDelay): } continue } @@ -1158,16 +1160,18 @@ func (h *OpenAIGatewayHandler) Messages(c *gin.Context) { retryLimit := account.GetPoolModeRetryCount() if sameAccountRetryCount[account.ID] < retryLimit { sameAccountRetryCount[account.ID]++ + retryDelay := sameAccountRetryDelayFor(failoverErr, sameAccountRetryCount[account.ID]) reqLog.Warn("openai_messages.pool_mode_same_account_retry", zap.Int64("account_id", account.ID), zap.Int("upstream_status", failoverErr.StatusCode), zap.Int("retry_limit", retryLimit), zap.Int("retry_count", sameAccountRetryCount[account.ID]), + zap.Duration("retry_delay", retryDelay), ) select { case <-c.Request.Context().Done(): return - case <-time.After(sameAccountRetryDelay): + case <-time.After(retryDelay): } continue }