xzj7019 commented on code in PR #20642:
URL: https://github.com/apache/doris/pull/20642#discussion_r1226268711
##########
fe/fe-core/src/main/java/org/apache/doris/nereids/cost/CostModelV1.java:
##########
@@ -191,7 +191,7 @@ public Cost visitPhysicalDistribute(
return CostV1.of(
0,
0,
- childStatistics.getRowCount() * Math.pow(beNumber, 0.5));
Review Comment:
In cost model v1, we use row count as the unified metrics to measure the
cost. So here we would better to also use row-count based cost system, if we
need to distinguish the distribute cost for different cases, we can use a
factor parameter using dataSize as the input with a minimal value 1.
##########
fe/fe-core/src/main/java/org/apache/doris/nereids/cost/CostModelV1.java:
##########
@@ -161,27 +161,27 @@ public Cost
visitPhysicalPartitionTopN(PhysicalPartitionTopN<? extends Plan> par
@Override
public Cost visitPhysicalDistribute(
PhysicalDistribute<? extends Plan> distribute, PlanContext
context) {
+ int kBytes = 1024;
Statistics childStatistics = context.getChildStatistics(0);
DistributionSpec spec = distribute.getDistributionSpec();
+ int beNumber =
ConnectContext.get().getEnv().getClusterInfo().getBackendsNumber(true);
+ beNumber = Math.max(1, beNumber);
+ double dataSize = childStatistics.computeSize() / kBytes; // in K bytes
// shuffle
if (spec instanceof DistributionSpecHash) {
return CostV1.of(
0,
0,
- childStatistics.getRowCount());
Review Comment:
In cost model v1, we use row count as the unified metrics to measure the
cost. So here we would better to also use row-count based cost system, if we
need to distinguish the distribute cost for different cases, we can use a
factor parameter using dataSize as the input with a minimal value 1.
--
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]
