-
Notifications
You must be signed in to change notification settings - Fork 1.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Push down filter plan for non-unnest column #11019
Conversation
Signed-off-by: jayzhan211 <[email protected]>
Signed-off-by: jayzhan211 <[email protected]>
Signed-off-by: jayzhan211 <[email protected]>
Signed-off-by: jayzhan211 <[email protected]>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Thank you very much @jayzhan211 -- looks great to me
I also updated the ticket description to reflect that this PR would close #11016 as far as I understand
cc @jonahgao
@@ -73,11 +73,29 @@ explain select uc2, column1 from (select unnest(column2) as uc2, column1 from v | |||
---- | |||
logical_plan | |||
01)Projection: unnest(v.column2) AS uc2, v.column1 | |||
02)--Filter: unnest(v.column2) > Int64(3) AND v.column1 = Int64(2) | |||
02)--Filter: unnest(v.column2) > Int64(3) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you please also update the comment a few lines above to reflect the issue is fixed?
# Could push the filter (column1 = 2) down below unnest
# https://github.com/apache/datafusion/issues/11016
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Looks good to me!
I am considering whether it is beneficial to wait for the next rewrite pass to handle the unnest input, because we already have the capability to push down filters below projections. This might simplify the implementation, and we don't need to assume that the unnest input is a projection.
LogicalPlan::Unnest(mut unnest) => {
let predicates = split_conjunction_owned(filter.predicate.clone());
let mut non_unnest_predicates = vec![];
let mut unnest_predicates = vec![];
for predicate in predicates {
// collect all the Expr::Column in predicate recursively
let mut accum: HashSet<Column> = HashSet::new();
expr_to_columns(&predicate, &mut accum)?;
if unnest.exec_columns.iter().any(|c| accum.contains(c)) {
unnest_predicates.push(predicate);
} else {
non_unnest_predicates.push(predicate);
}
}
// Unnest predicates should not be pushed down.
// If no non-unnest predicates exist, early return
if non_unnest_predicates.is_empty() {
filter.input = Arc::new(LogicalPlan::Unnest(unnest));
return Ok(Transformed::no(LogicalPlan::Filter(filter)));
}
let new_filter = LogicalPlan::Filter(Filter::try_new(
conjunction(non_unnest_predicates).unwrap(),
unnest.input.clone(),
)?);
unnest.input = Arc::new(new_filter);
if unnest_predicates.is_empty() {
Ok(Transformed::yes(LogicalPlan::Unnest(unnest)))
} else {
filter.predicate = conjunction(unnest_predicates).unwrap();
filter.input = Arc::new(LogicalPlan::Unnest(unnest));
Ok(Transformed::no(LogicalPlan::Filter(filter)))
}
}
I see. It might worth to try. |
Signed-off-by: jayzhan211 <[email protected]>
Signed-off-by: jayzhan211 <[email protected]>
)?); | ||
|
||
// try push down recursively | ||
let new_plan = self.rewrite(filter_with_unnest_input, _config)?; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think the optimizer already handles recursion, so it might not be necessary to call rewrite here.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Agree!
// Filter | ||
// Unenst Input (Projection) | ||
|
||
let unnest_input = std::mem::take(&mut unnest.input); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
How about directly cloning the input since it is an Arc
? This way, we don't need to construct a Default LogicalPlan either.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We have a plan to convert Arc<LogicalPlan>
to Box<LogicalPlan>
, but before that we need to make sure there is no clone exist, so I think we should try our best to avoid clone. It is also the reason why we have rewrite
in Logical optimizer rule that takes LogicalPlan
instead of Arc<LogicalPlan>
Ref: #9637 (comment)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Since there's a purpose for doing this, make sense to me. Thanks @jayzhan211
Signed-off-by: jayzhan211 <[email protected]>
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM! Thanks @jayzhan211
@@ -281,6 +281,15 @@ pub enum LogicalPlan { | |||
RecursiveQuery(RecursiveQuery), | |||
} | |||
|
|||
impl Default for LogicalPlan { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
👍
* push down non-unnest only Signed-off-by: jayzhan211 <[email protected]> * add doc Signed-off-by: jayzhan211 <[email protected]> * add doc Signed-off-by: jayzhan211 <[email protected]> * cleanup Signed-off-by: jayzhan211 <[email protected]> * rewrite unnest push donw filter Signed-off-by: jayzhan211 <[email protected]> * remove comment Signed-off-by: jayzhan211 <[email protected]> * avoid double recurisve Signed-off-by: jayzhan211 <[email protected]> --------- Signed-off-by: jayzhan211 <[email protected]>
* push down non-unnest only Signed-off-by: jayzhan211 <[email protected]> * add doc Signed-off-by: jayzhan211 <[email protected]> * add doc Signed-off-by: jayzhan211 <[email protected]> * cleanup Signed-off-by: jayzhan211 <[email protected]> * rewrite unnest push donw filter Signed-off-by: jayzhan211 <[email protected]> * remove comment Signed-off-by: jayzhan211 <[email protected]> * avoid double recurisve Signed-off-by: jayzhan211 <[email protected]> --------- Signed-off-by: jayzhan211 <[email protected]>
* push down non-unnest only Signed-off-by: jayzhan211 <[email protected]> * add doc Signed-off-by: jayzhan211 <[email protected]> * add doc Signed-off-by: jayzhan211 <[email protected]> * cleanup Signed-off-by: jayzhan211 <[email protected]> * rewrite unnest push donw filter Signed-off-by: jayzhan211 <[email protected]> * remove comment Signed-off-by: jayzhan211 <[email protected]> * avoid double recurisve Signed-off-by: jayzhan211 <[email protected]> --------- Signed-off-by: jayzhan211 <[email protected]>
Which issue does this PR close?
Follow on #10991
Closes #11016
Rationale for this change
In #10991, if we see unnest column, we skip the push down optimization.
In this PR, we select the non-unnest columns to push down, even there is unnest columns besides.
What changes are included in this PR?
Are these changes tested?
Are there any user-facing changes?