[ https://issues.apache.org/jira/browse/HIVE-20090?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16535603#comment-16535603 ]
Gopal V commented on HIVE-20090: -------------------------------- [~jcamachorodriguez]: you end up calling {code} + GenTezUtils.removeBranch(p.getLeft()); {code} on the same item twice, which needs GenTezUtils.removeBranch to do this {code} diff --git a/ql/src/java/org/apache/hadoop/hive/ql/parse/GenTezUtils.java b/ql/src/java/org/apache/hadoop/hive/ql/parse/GenTezUtils.java index bb0de94..2aae6dc 100644 --- a/ql/src/java/org/apache/hadoop/hive/ql/parse/GenTezUtils.java +++ b/ql/src/java/org/apache/hadoop/hive/ql/parse/GenTezUtils.java @@ -496,6 +496,10 @@ public static void removeBranch(Operator<?> event) { while (curr.getChildOperators().size() <= 1) { child = curr; + if (curr.getParentOperators().isEmpty()) { + // this is already a dead-branch + return; + } curr = curr.getParentOperators().get(0); } {code} > Extend creation of semijoin reduction filters to be able to discover new > opportunities > -------------------------------------------------------------------------------------- > > Key: HIVE-20090 > URL: https://issues.apache.org/jira/browse/HIVE-20090 > Project: Hive > Issue Type: Improvement > Components: Physical Optimizer > Reporter: Jesus Camacho Rodriguez > Assignee: Jesus Camacho Rodriguez > Priority: Major > Attachments: HIVE-20090.01.patch, HIVE-20090.02.patch > > > Assume the following plan: > {noformat} > TS[0] - RS[1] - JOIN[4] - RS[5] - JOIN[8] - FS[9] > TS[2] - RS[3] - JOIN[4] > TS[6] - RS[7] - JOIN[8] > {noformat} > Currently, {{TS\[6\]}} may only be reduced with the output of {{RS\[5\]}}, > i.e., input to join between both subplans. > However, it may be useful to consider other possibilities too, e.g., reduced > by the output of {{RS\[1\]}} or {{RS\[3\]}}. For instance, this is important > when, given a large plan, an edge between {{RS[5]}} and {{TS[0]}} would > create a cycle, while an edge between {{RS[1]}} and {{TS[6]}} would not. > This patch comprises two parts. First, it creates additional predicates when > possible. Secondly, it removes duplicate semijoin reduction > branches/predicates, e.g., if another semijoin that consumes the output of > the same expression already reduces a certain table scan operator (heuristic, > since this may not result in most efficient plan in all cases). Ultimately, > the decision on whether to use one or another should be cost-driven > (follow-up). -- This message was sent by Atlassian JIRA (v7.6.3#76005)