Sorry for the misktake, correct it. "but I'm afraid that we have no other better way to avoid it.
Best regards, Yuxia ----- 原始邮件 ----- 发件人: "yuxia" <luoyu...@alumni.sjtu.edu.cn> 收件人: "dev" <dev@flink.apache.org> 发送时间: 星期一, 2023年 2 月 13日 上午 10:11:53 主题: Re: Confusion about some overlapping functionality of SupportsProjectionPushDown and SupportsReadingMetadata Hi, Ran Tao. I agree to Shengkai that physical column pushdown and metadata processing are different things and should be processed separately. To me, it'll be more weird that SupportsReadingMetadata do the both. I agree you that that the developer may have unexpected behaviors due to the sequence problem, but I don't think we have no other better way to avoid it. Best regards, Yuxia ----- 原始邮件 ----- 发件人: "fskmine" <fskm...@gmail.com> 收件人: "dev" <dev@flink.apache.org> 发送时间: 星期五, 2023年 2 月 10日 下午 12:02:36 主题: Re: Confusion about some overlapping functionality of SupportsProjectionPushDown and SupportsReadingMetadata Hi, Ran. I think it's a little difficult to split the rule into two parts. Because the ProjectDown and ReadingMetadata both need to reorder the fields. The ReadingMetadata requires the metadata columns to be at the last and the ProjectPushDown now is responsible to reorder the columns as the user specified. One more concern, the push-down optimization occurs in the logical phase that uses the volcano planner. I am not sure whether the rule will be applied at last because metadata push-down doesn't change the cost. Best, Shengkai
