Indeed this PR only addresses the "NONE 1st" case. As for the "NONE 2nd" case (following a 1st OK_NEW_SCHEMA) - this would open up the can of worms we saw before - early sniffing the probe side cranks the actual execution there (e.g., another Hash Join below an Exchange) while other parts of the query are still in schema discovery.
Not sure if in all cases where an operator finds it has no data to return during schema discovery, a NONE is returned first. Note that the NONE does have a schema (e.g., so a Right Join can return nulls for those missing columns). Maybe we can increase the impact of this PR by checking all these cases, and documenting/testing that our protocol allows for a NONE to be returned first (i.e. during schema discovery). [ Full content available at: https://github.com/apache/drill/pull/1480 ] This message was relayed via gitbox.apache.org for [email protected]
