[
https://issues.apache.org/jira/browse/HIVE-21690?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16834241#comment-16834241
]
Jesus Camacho Rodriguez edited comment on HIVE-21690 at 5/6/19 10:57 PM:
-------------------------------------------------------------------------
{quote}
One approach to fix this is to localize the cost computation to the rule
itself, i.e compute the non-cumulative cost of existing aggregate and join and
compare it with new cost of new aggregates, join and top aggregate.
Better approach in my opinion would be to fix the cost model and take aggregate
cost into account (along with the join). This could affect other queries and
can cause performance regression but those will most likely be issues with the
planning and should be investigated and fixed.
{quote}
In principle, second approach seems the logical choice since it takes the cost
model closer to execution actual cost. However, it can easily backfire with
current implementation of join reordering costing, which only considers cost
for join operation and builds on that assumption. Why should we consider only
aggregate operators? What about other operators?
Before pushing such a change, I would argue that we need further evaluation on
how this will affect join reordering and the regressions that we will get wrt
previous benchmarks.
Since cost model is pluggable, have you thought about creating a cost model
that extends the join reordering (default) one with cost calculation for the
Aggregate operator? You could use the new cost model when you trigger this
rule. In a follow-up, you can study whether using the same cost model for join
reordering makes sense or not, and evaluate the merit of that change for join
reordering on its own.
was (Author: jcamachorodriguez):
{quote}
One approach to fix this is to localize the cost computation to the rule
itself, i.e compute the non-cumulative cost of existing aggregate and join and
compare it with new cost of new aggregates, join and top aggregate.
Better approach in my opinion would be to fix the cost model and take aggregate
cost into account (along with the join). This could affect other queries and
can cause performance regression but those will most likely be issues with the
planning and should be investigated and fixed.
{quote}
In principle, second approach seems the logical choice since it takes the cost
model closer to execution actual cost. However, it can easily backfire with
current implementation of join reordering costing, which only considers cost
for join operation and builds on that assumption. Why should we consider only
aggregate operators? What about other operators?
Before pushing such a change, I would argue that we need further evaluation on
how this will affect join reordering and the regressions that we will get wrt
previous benchmarks.
Since cost model is pluggable, have you thought about creating a cost model
that extends the join reordering (default) one with cost calculation for the
Aggregate operator? You could use the new cost model when you trigger this
rule. In a follow-up, you can study whether using the same cost model for join
reordering makes sense or not, and evaluate the merit of that change for join
reordering on its own.
> Support outer joins with HiveAggregateJoinTransposeRule and turn it on by
> default
> ---------------------------------------------------------------------------------
>
> Key: HIVE-21690
> URL: https://issues.apache.org/jira/browse/HIVE-21690
> Project: Hive
> Issue Type: Improvement
> Components: Query Planning
> Reporter: Vineet Garg
> Assignee: Vineet Garg
> Priority: Major
> Attachments: HIVE-21690.1.patch
>
>
> 1) This optimization is off by default. We would like to turn on this
> optimization wherein group by is pushed down to join, in some cases top
> aggregate is removed but in most of the cases this optimization adds extra
> aggregate nodes. To measure if those extra aggregates are beneficial or not
> (they might add extra overhead without reducing rows) cost is computed and
> compared b/w previous plan and new plan.
> Since Hive's cost model only consider JOIN's cost and discard cost of rest of
> the nodes, this comparison always favor new plan (since adding aggregate
> beneath join reduces the total number of rows processed by the join and
> therefore reduces the join cost). Therefore turning on this optimization with
> existing cost model is not a good idea.
> One approach to fix this is to localize the cost computation to the rule
> itself, i.e compute the non-cumulative cost of existing aggregate and join
> and compare it with new cost of new aggregates, join and top aggregate.
> Better approach in my opinion would be to fix the cost model and take
> aggregate cost into account (along with the join). This could affect other
> queries and can cause performance regression but those will most likely be
> issues with the planning and should be investigated and fixed.
> 2) This optimization currently only support INNER JOIN. This can be extended
> to support OUTER joins.
>
> cc [~jcamachorodriguez] [~ashutoshc] [~gopalv]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
