englefly commented on code in PR #27259:
URL: https://github.com/apache/doris/pull/27259#discussion_r1399933480
##########
fe/fe-core/src/main/java/org/apache/doris/nereids/cost/CostModelV1.java:
##########
@@ -299,12 +303,41 @@ public Cost visitPhysicalHashJoin(
0
);
}
- return CostV1.of(context.getSessionVariable(), leftRowCount +
rightRowCount + outputRowCount,
+ return CostV1.of(context.getSessionVariable(),
+ (leftRowCount + rightRowCount + outputRowCount)
+ * skewPenalty(physicalHashJoin, probeStats,
buildStats, totalInstanceNumber),
rightRowCount,
0
);
}
+ private double skewPenalty(PhysicalHashJoin<? extends Plan, ? extends
Plan> physicalHashJoin,
+ Statistics probeStats,
+ Statistics buildStats,
+ int totalInstanceNumber) {
+ double penalty = 1.0;
+ for (Expression conj : physicalHashJoin.getHashJoinConjuncts()) {
+ EqualTo eq = (EqualTo) JoinUtils.swapEqualToForChildrenOrder(
+ (EqualTo) conj, physicalHashJoin.left().getOutputSet());
+ ColumnStatistic leftColStats =
probeStats.findColumnStatistics(eq.left());
+ if (leftColStats != null && !leftColStats.isUnKnown()) {
+ if (leftColStats.ndv < beNumber) {
+ penalty = totalInstanceNumber / Math.max(1.0,
leftColStats.ndv);
+ break;
+ }
+ }
+
+ ColumnStatistic rightColStats =
buildStats.findColumnStatistics(eq.right());
+ if (rightColStats != null && !rightColStats.isUnKnown()) {
+ if (rightColStats.ndv < beNumber) {
+ penalty = totalInstanceNumber / Math.max(1.0,
leftColStats.ndv);
Review Comment:
if right.ndv=1, then only one instance build hash table with full data, and
others build empty hash table. All the workload is concentrated on a single
node.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
