optimize archive retry claim SQL

internal/service/taskcenter/archive.go

@@ -113,16 +113,32 @@ func (s *Service) claimArchiveRows(ctx context.Context, batchSize int) ([]model.
 			rows = append(rows, staleRows...)
 		}
 		if remaining := batchSize - len(rows); remaining > 0 {
-			var retryableRows []model.TaskCenterEventArchiveOutbox
-			// PENDING/FAILED 只按状态和创建时间抢占，避免和 PROCESSING 超时接管揉成 OR 后放大扫描。
-			if err := tx.Clauses(clause.Locking{Strength: "UPDATE", Options: "SKIP LOCKED"}).
-				Where("status IN ?", []string{ArchiveStatusPending, ArchiveStatusFailed}).
-				Order("create_time asc, id asc").
-				Limit(remaining).
-				Find(&retryableRows).Error; err != nil {
-				return err
+			pendingLimit := taskCenterArchivePendingClaimLimit(remaining)
+			for _, retryable := range []struct {
+				status string
+				limit  int
+			}{
+				{status: ArchiveStatusPending, limit: pendingLimit},
+				{status: ArchiveStatusFailed, limit: remaining},
+			} {
+				if retryable.limit <= 0 || remaining <= 0 {
+					continue
+				}
+				if retryable.limit > remaining {
+					retryable.limit = remaining
+				}
+				var retryableRows []model.TaskCenterEventArchiveOutbox
+				// PENDING 和 FAILED 分开抢占，避免 IN 跨 status 排序触发 filesort，同时给 FAILED 保留重试额度。
+				if err := tx.Clauses(clause.Locking{Strength: "UPDATE", Options: "SKIP LOCKED"}).
+					Where("status = ?", retryable.status).
+					Order("create_time asc, id asc").
+					Limit(retryable.limit).
+					Find(&retryableRows).Error; err != nil {
+					return err
+				}
+				rows = append(rows, retryableRows...)
+				remaining -= len(retryableRows)
 			}
-			rows = append(rows, retryableRows...)
 		}
 		if len(rows) == 0 {
 			return nil
@@ -164,6 +180,20 @@ func taskCenterArchiveClaimLimits(batchSize int) (int, int) {
 	return staleLimit, batchSize - staleLimit
 }
 
+func taskCenterArchivePendingClaimLimit(retryableLimit int) int {
+	if retryableLimit <= 1 {
+		return retryableLimit
+	}
+	failedReserve := retryableLimit / 10
+	if failedReserve < 1 {
+		failedReserve = 1
+	}
+	if failedReserve > 100 {
+		failedReserve = 100
+	}
+	return retryableLimit - failedReserve
+}
+
 func (s *Service) newCOSClient() (*cos.Client, error) {
 	bucketURL, err := url.Parse(fmt.Sprintf("https://%s.cos.%s.myqcloud.com",
 		strings.TrimSpace(s.cfg.TaskCenter.Archive.Bucket),

internal/service/taskcenter/taskcenter_test.go

@@ -87,6 +87,28 @@ func TestTaskCenterArchiveClaimLimitsReserveStaleRows(t *testing.T) {
 	}
 }
 
+func TestTaskCenterArchivePendingClaimLimitReservesFailedRows(t *testing.T) {
+	tests := []struct {
+		name           string
+		retryableLimit int
+		want           int
+	}{
+		{name: "empty", retryableLimit: 0, want: 0},
+		{name: "single row stays pending first", retryableLimit: 1, want: 1},
+		{name: "small retry batch keeps one failed slot", retryableLimit: 4, want: 3},
+		{name: "normal retry batch reserves ten percent", retryableLimit: 450, want: 405},
+		{name: "large retry batch caps failed reserve", retryableLimit: 900, want: 810},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := taskCenterArchivePendingClaimLimit(tt.retryableLimit); got != tt.want {
+				t.Fatalf("taskCenterArchivePendingClaimLimit(%d) = %d, want %d", tt.retryableLimit, got, tt.want)
+			}
+		})
+	}
+}
+
 func TestDailyTaskConfigEventListAndClaim(t *testing.T) {
 	java := &taskCenterTestGateway{events: map[string]bool{}}
 	service, db := newTestService(t, java)