Hi camelia, I'm sorry for late response. I've just came back home from the family meeting. I leave in-line comments on your question.
Best regards, Hyunsik On Sep 7, 2013, at 8:42 PM, camelia c <[email protected]> wrote: > Hello, > > I resend You an updated list of questions that I have. For some of the > ancient ones, I found the answer already. > > 1) In MergeJoinExec, what is the purpose of the innerTupleSlots and > outerTupleSlots and can You please give me an example of how they are filled, > based on a dummy data set ? Merge join forwards each relation in order to find the same join key tuples. Each of them keeps a list of tuples whose join keys are same. Consider the below examples where there are two relations to be joined and the first column of each relation is the join key. ----------------------------------- Two relations to be joined ----------------------------------- Left Right (1, A) (1, B) (1, C) (1, C) (3, D) (1, D) (2, E) MergeJoin first finds all the same key tuples for each relation. So, each tuple slot contains as follows: outerTupleSlots : (1, A), (1,C) innerTupleSlots : (1,B), (1, C), (1,D) Then, MergeJoin leads to joined tuples. In the above example, MergeJoin results in 6 tuples (2 x 3). > > 2) I understood from a talk that the MergeJoinExec has some issues and that > Mr Jihoon is trying to fix them. Can I rely on the current version of > MergeJoinExec to extend it for FullOuter_MergeJoinExec and > RightOuter_MergeJoinExec? MergeJoinExec does not have any problem. It is correct. There was a misunderstood. > > 3) Given a JoinNode anywhere in the logical query plan, how can we obtain the > block name containing it? > Even for a single-block query, how do we find for a JoinNode that it belongs > to @ROOT, for example? > > More precisely, in class OuterJoinRewriteRule, in method > public LogicalNode visitJoin(LogicalPlan plan, JoinNode joinNode, > Stack<LogicalNode> stack, Integer depth) > > I tried to do > plan.getBlock(joinNode).getName() > but I receive a Null Pointer Exception. > The current API cannot what you want. The API needs to be improved for supporting that. Probably, that is archived by modifying BasicLogicalNodeVisitor's visitChild method to call visitXXXNode method with some object including a current block name. I'll create a jira issue for this improvement. > > > I look forward to receiving Your answer! > > Yours sincerely, > Camelia
