[
https://issues.apache.org/jira/browse/IGNITE-20383?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Aleksey Plekhanov updated IGNITE-20383:
---------------------------------------
Description:
Sometimes, if join inputs are not collocated it's worth to broadcast one of the
inputs, for example, query:
{code:sql}
SELECT * FROM emps WHERE emps.salary = (SELECT AVG(emps.salary) FROM emps){code}
Currently has plan:
{noformat}
IgniteProject(ID=[$0], NAME=[$1], SALARY=[$2])
IgniteNestedLoopJoin(condition=[=($2, $3)], joinType=[inner])
IgniteExchange(distribution=[single])
IgniteTableScan(table=[[PUBLIC, EMPS]])
IgniteReduceHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
IgniteExchange(distribution=[single])
IgniteMapHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[TST],
requiredColumns=[{2}], collation=[[2 ASC-nulls-first]])
{noformat}
But this plan is not optimal, since we should send entire table EMP from all
nodes to the single node. For such a query it's better to broadcast result of
the aggregation, in this case plan will be something like:
{noformat}
IgniteExchange(distribution=[single])
IgniteProject(...)
IgniteCorrelatedNestedLoopJoin(...)
IgniteExchange(distribution=[broadcast])
IgniteReduceHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
IgniteExchange(distribution=[single])
IgniteMapHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[SALARY_IDX])
IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[SALARY_IDX])
{noformat}
But currently we don't try to convert any of the join inputs to the broadcast
distribution. We should try to do this.
was:
Sometimes, if join inputs are not collocated it's worth to broadcast one of the
inputs, for example, query:
{code:sql}
SELECT * FROM emps WHERE emps.salary = (SELECT AVG(emps.salary) FROM emps){code}
Currently has plan:
{noformat}
IgniteProject(ID=[$0], NAME=[$1], SALARY=[$2])
IgniteNestedLoopJoin(condition=[=($2, $3)], joinType=[inner])
IgniteExchange(distribution=[single])
IgniteTableScan(table=[[PUBLIC, EMPS]])
IgniteReduceHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
IgniteExchange(distribution=[single])
IgniteMapHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[TST],
requiredColumns=[{2}], collation=[[2 ASC-nulls-first]])
{noformat}
But this plan is not optimal, since we should send entire table EMP from all
nodes to the single node. For such a query it's better to broadcast result of
the aggregation, in this case plan will be something like:
{noformat}
IgniteExchange(distribution=[single])
IgniteProject(ID=[$0], NAME=[$1], SALARY=[$2])
IgniteCorrelatedNestedLoopJoin(...)
IgniteExchange(distribution=[broadcast])
IgniteReduceHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
IgniteExchange(distribution=[single])
IgniteMapHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[SALARY_IDX])
IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[SALARY_IDX])
{noformat}
But currently we don't try to convert any of the join inputs to the broadcast
distribution. We should try to do this.
> Calcite engine. Convert one input of a join to the broadcast distribution
> -------------------------------------------------------------------------
>
> Key: IGNITE-20383
> URL: https://issues.apache.org/jira/browse/IGNITE-20383
> Project: Ignite
> Issue Type: Improvement
> Reporter: Aleksey Plekhanov
> Assignee: Aleksey Plekhanov
> Priority: Major
> Labels: calcite, ise
>
> Sometimes, if join inputs are not collocated it's worth to broadcast one of
> the inputs, for example, query:
> {code:sql}
> SELECT * FROM emps WHERE emps.salary = (SELECT AVG(emps.salary) FROM
> emps){code}
> Currently has plan:
> {noformat}
> IgniteProject(ID=[$0], NAME=[$1], SALARY=[$2])
> IgniteNestedLoopJoin(condition=[=($2, $3)], joinType=[inner])
> IgniteExchange(distribution=[single])
> IgniteTableScan(table=[[PUBLIC, EMPS]])
> IgniteReduceHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
> IgniteExchange(distribution=[single])
> IgniteMapHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
> IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[TST],
> requiredColumns=[{2}], collation=[[2 ASC-nulls-first]])
> {noformat}
> But this plan is not optimal, since we should send entire table EMP from all
> nodes to the single node. For such a query it's better to broadcast result of
> the aggregation, in this case plan will be something like:
> {noformat}
> IgniteExchange(distribution=[single])
> IgniteProject(...)
> IgniteCorrelatedNestedLoopJoin(...)
> IgniteExchange(distribution=[broadcast])
> IgniteReduceHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
> IgniteExchange(distribution=[single])
> IgniteMapHashAggregate(group=[{}], AVG(EMPS.SALARY)=[AVG($0)])
> IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[SALARY_IDX])
> IgniteIndexScan(table=[[PUBLIC, EMPS]], index=[SALARY_IDX])
> {noformat}
> But currently we don't try to convert any of the join inputs to the broadcast
> distribution. We should try to do this.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
