Github user cloud-fan commented on the issue: https://github.com/apache/spark/pull/21143 If we don't care about whole stage codegen, I think it's possible to dynamically change the filter condition at executors and codegen it. Now the problem becomes, how would the data source API look like? I'd image some systems have to do filter pushdown at the driver side, so that they can connect to the metadata service only once. One possible way is to allow users to mixin `SupportsPushDownFilter` to both `DataSourceReader` and `DataReaderFactory`.
--- --------------------------------------------------------------------- To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org For additional commands, e-mail: reviews-h...@spark.apache.org