@sohami I agree that guarantee was not provided before. However, I am proposing to change the contract in the way described in the comment. Note no data would be associated with a record batch after it returns NONE. It would only have a VectorContainer with **EMPTY** columns corresponding to the last seen valid schema, and return a record count of 0. In practice almost all operators already do this (except they don't always zero their container).
The advantage of adding this new guarantee is that downstream operators do not have to do book keeping when the upstream can do it for us very easily. The details of why this was done are in the JIRA. There were many ways to fix the issue in HashJoin but this is by far the cleanest, and less risky since most of the code in HashJoin has no unit tests. It also has the advantage of clarifying the behavior of RecordBatches for future operator implementations. So I felt this approach was a win win and the best way to go about fixing the issue. However, there should be a follow up Jira to validate the new contract I am proposing for the rest of the operators, not just unordered reciever. [ Full content available at: https://github.com/apache/drill/pull/1472 ] This message was relayed via gitbox.apache.org for [email protected]
