[
https://issues.apache.org/jira/browse/IGNITE-21948?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Andrey Mashenkov updated IGNITE-21948:
--------------------------------------
Description:
Motivation.
We convert LogicalAggregate to map-reduce unconditionally (if possible), which
blows plan search space.
Some plans, which have reduce-phase followed by map-phase without Exchange,
make no sense and will be thrown away after applying cost model (if no - there
is a bug).
Suggestion.
What if we will emit only plans with colocated aggregates?
We can write rules to pushdown colocated aggregate through Exchange via
splitting into map-reduce phases if colocation can't be preserved.
If Exchange can preserve colocation, then aggregate is pushed gown without
splitting.
Let's investigate, if we can eliminate (or significantly decrease number of
plans)
# with Exchange-less Reduce->Map
# map-reduce aggregates on colocated data.
was:
Motivation.
We convert LogicalAggregate to map-reduce unconditionally (if possible), which
blows plan search space.
Some plans, which have reduce-phase followed by map-phase without Exchange,
make no sense and will be thrown away after applying cost model (if no - there
is a bug).
Suggestion.
What if we will emit only plans with colocated aggregates?
We can write rules to pushdown colocated aggregate through Exchange via
splitting into map-reduce phases if colocation can't be preserved.
If Exchange can preserve colocation, then aggregate is pushed gown without
splitting.
Let's investigate, if we can eliminate (or significantly decrease number of
plans)
# with Exchange-less Reduce->Map
# map-reduce aggregates on colocated data.
> Rethink distributed aggregates converter rules.
> -----------------------------------------------
>
> Key: IGNITE-21948
> URL: https://issues.apache.org/jira/browse/IGNITE-21948
> Project: Ignite
> Issue Type: Improvement
> Components: sql
> Reporter: Andrey Mashenkov
> Priority: Major
> Labels: ignite-3, perfomance
>
> Motivation.
> We convert LogicalAggregate to map-reduce unconditionally (if possible),
> which blows plan search space.
> Some plans, which have reduce-phase followed by map-phase without Exchange,
> make no sense and will be thrown away after applying cost model (if no -
> there is a bug).
> Suggestion.
> What if we will emit only plans with colocated aggregates?
> We can write rules to pushdown colocated aggregate through Exchange via
> splitting into map-reduce phases if colocation can't be preserved.
> If Exchange can preserve colocation, then aggregate is pushed gown without
> splitting.
>
> Let's investigate, if we can eliminate (or significantly decrease number of
> plans)
> # with Exchange-less Reduce->Map
> # map-reduce aggregates on colocated data.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
