Aleksandr Efimov has posted comments on this change. ( http://gerrit.cloudera.org:8080/24592 )
Change subject: IMPALA-15189: Support HBO for SortNode cardinality ...................................................................... Patch Set 6: (1 comment) http://gerrit.cloudera.org:8080/#/c/24592/6/fe/src/main/java/org/apache/impala/planner/DistributedPlanner.java File fe/src/main/java/org/apache/impala/planner/DistributedPlanner.java: http://gerrit.cloudera.org:8080/#/c/24592/6/fe/src/main/java/org/apache/impala/planner/DistributedPlanner.java@1352 PS6, Line 1352: lowerTopN.setTopNMergeParent(upperTopN); > `lowerTopN` has already run `init()` and `computeStats()` before this split I'd soften "harmless" here. ExchangeNode copies the local Top-N cardinality and divides it by getNumNodes() for the queue memory estimate, so it uses it as a total. With two senders each returning five rows for the same ten groups and no ties, 100 rows cross the exchange while the final Top-N returns 50. My suggestion to recompute doesn't solve this either: the non-HBO estimate can have the same problem. Could we document this approximation and track the exchange costing separately? -- To view, visit http://gerrit.cloudera.org:8080/24592 To unsubscribe, visit http://gerrit.cloudera.org:8080/settings Gerrit-Project: Impala-ASF Gerrit-Branch: master Gerrit-MessageType: comment Gerrit-Change-Id: Ib829887a91593bee124d56e661e714575fe3be97 Gerrit-Change-Number: 24592 Gerrit-PatchSet: 6 Gerrit-Owner: Quanlong Huang <[email protected]> Gerrit-Reviewer: Aleksandr Efimov <[email protected]> Gerrit-Reviewer: Impala Public Jenkins <[email protected]> Gerrit-Reviewer: Quanlong Huang <[email protected]> Gerrit-Comment-Date: Sun, 06 Sep 2026 07:15:27 +0000 Gerrit-HasComments: Yes
