ORM易写但易慢，关键在查询数与返回量。首先使用select_related优化一对一/外键，prefetch_related优化一对多/多对多；其次只取所需字段values/only/defer降低IO；再次用exists判断存在性，不要count全表；聚合统计使用annotate避免Python层循环。必要时加索引并关注迁移顺序。示例：

qs = (Book.objects
      .select_related("publisher")
      .prefetch_related("authors")
      .only("id","title","publisher__name"))