Skip to content

Collapse generate_series when output columns are projected away #22753

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.

Sign up for GitHub

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

Closed
frankmcsherry wants to merge 3 commits into from
+134 −49
Closed

Collapse generate_series when output columns are projected away #22753

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
059a0cb
Collapse generate_series when output columns are projected away
frankmcsherry Oct 27, 2023
f505629
Cast i32 to i64 before repeat
frankmcsherry Oct 30, 2023
2a956c0
Optimize fold_reduce_constant
frankmcsherry Nov 7, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
97 changes: 48 additions & 49 deletions src/transform/src/fold_constants.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -432,10 +432,10 @@ impl FoldConstants {
rows: &[(Row, Diff)],
limit: Option<usize>,
) -> Option<Result<Vec<(Row, Diff)>, EvalError>> {
// Build a map from `group_key` to `Vec<Vec<an, ..., a1>>)`,
// where `an` is the input to the nth aggregate function in
// `aggregates`.
let mut groups = BTreeMap::new();
// A list of all keys, independent of the aggregates.
let mut keys = BTreeSet::new();
// For each aggregate, a map from keys to a list of (val, diff) pairs.
let mut groups = vec![BTreeMap::new(); aggregates.len()];
let temp_storage2 = RowArena::new();
let mut row_buf = Row::default();
let mut limit_remaining = limit.unwrap_or(usize::MAX);
Expand All @@ -451,10 +451,10 @@ impl FoldConstants {
))));
}

if limit_remaining < *diff as usize {
if limit_remaining < 1 as usize {
return None;
}
limit_remaining -= *diff as usize;
limit_remaining -= 1 as usize;
Comment on lines +454 to +457
Copy link
Copy Markdown
Member

@antiguru antiguru Oct 27, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks like this doesn't need a cast.


let datums = row.unpack();
let temp_storage = RowArena::new();
Expand All @@ -463,25 +463,23 @@ impl FoldConstants {
.map(|e| e.eval(&datums, &temp_storage2))
.collect::<Result<Vec<_>, _>>()
{
Ok(key) => key,
Ok(key) => std::rc::Rc::new(key),
Err(e) => return Some(Err(e)),
};
let val = match aggregates
.iter()
.map(|agg| {
row_buf
.packer()
.extend([agg.expr.eval(&datums, &temp_storage)?]);
Ok::<_, EvalError>(row_buf.clone())
})
.collect::<Result<Vec<_>, _>>()
{
Ok(val) => val,
Err(e) => return Some(Err(e)),
};
let entry = groups.entry(key).or_insert_with(Vec::new);
for _ in 0..*diff {
entry.push(val.clone());
keys.insert((*key).clone());
for (aggregate, group) in aggregates.iter().zip(groups.iter_mut()) {
let entry = group.entry(key.clone()).or_insert_with(Vec::new);
row_buf
.packer()
.extend([aggregate.expr.eval(&datums, &temp_storage).ok()?]);
entry.push((row_buf.clone(), *diff));
}
}

// Consolidate each group, in the event that there are like records.
for group in groups.iter_mut() {
for vals in group.values_mut() {
differential_dataflow::consolidation::consolidate(vals);
}
}

Expand All @@ -490,33 +488,34 @@ impl FoldConstants {
// `Vec<Vec<k1, ..., kn, r1, ..., rn>>`
// where kn is the nth column of the key and rn is the
// result of the nth aggregate function for that group.
let new_rows = groups
.into_iter()
.map({
let mut row_buf = Row::default();
move |(key, vals)| {
let temp_storage = RowArena::new();
row_buf.packer().extend(key.into_iter().chain(
aggregates.iter().enumerate().map(|(i, agg)| {
if agg.distinct {
agg.func.eval(
vals.iter()
.map(|val| val[i].unpack_first())
.collect::<BTreeSet<_>>(),
&temp_storage,
)
} else {
agg.func.eval(
vals.iter().map(|val| val[i].unpack_first()),
&temp_storage,
)
}
}),
));
(row_buf.clone(), 1)
}
let mut row_buf = Row::default();
let new_rows = keys
.iter()
.map(|key| {
let temp_storage = RowArena::new();
row_buf.packer().extend(key.into_iter().cloned().chain(
groups.iter_mut().zip(aggregates).map(|(group, aggr)| {
let values = &group[key];
if aggr.distinct {
aggr.func.eval(
values.iter().map(|(val, _count)| val.unpack_first()),
&temp_storage,
)
} else {
aggr.func.eval(
values.iter().flat_map(|(val, count)| {
std::iter::repeat(val.unpack_first())
.take((*count).try_into().unwrap())
}),
&temp_storage,
)
}
}),
));
(row_buf.clone(), 1)
})
.collect();
.collect::<Vec<_>>();

Some(Ok(new_rows))
}

Expand Down
86 changes: 86 additions & 0 deletions src/transform/src/movement/projection_pushdown.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -225,6 +225,92 @@ impl ProjectionPushdown {
for expr in exprs.iter() {
columns_to_pushdown.extend(expr.support());
}

// Check if any output columns of the flatmap are used, before removing them.
let output_columns_used = columns_to_pushdown.iter().any(|c| *c >= inner_arity);

// If we use none of the output columns, we can potentially simplify the flatmap.
// For example, `generate_series` could be simplified to `repeat`, plausibly.
// There may be other examples with known output counts, but I don't know them.
// The timestamp-based `generate_series` functions seemingly lack a division method.
if !output_columns_used {
use mz_expr::{BinaryFunc, TableFunc, UnaryFunc, VariadicFunc};

match func {
TableFunc::GenerateSeriesInt32 => {
// The arguments in `exprs` are `[start, stop, step]`.
// Per Brennan, the number of elements seems likely to be max (0, (stop - start + 1) / step)
let expr = MirScalarExpr::CallUnary {
func: UnaryFunc::CastInt32ToInt64(mz_expr::func::CastInt32ToInt64),
expr: Box::new(MirScalarExpr::CallVariadic {
func: VariadicFunc::Greatest,
exprs: vec![
MirScalarExpr::literal_ok(
mz_repr::Datum::Int32(0),
mz_repr::ScalarType::Int32,
),
MirScalarExpr::CallBinary {
func: BinaryFunc::DivInt32,
expr1: Box::new(MirScalarExpr::CallBinary {
func: BinaryFunc::SubInt32,
expr1: Box::new(MirScalarExpr::CallBinary {
func: BinaryFunc::AddInt32,
expr1: Box::new(exprs[1].clone()),
expr2: Box::new(MirScalarExpr::literal_ok(
mz_repr::Datum::Int32(1),
mz_repr::ScalarType::Int32,
)),
}),
expr2: Box::new(exprs[0].clone()),
}),
expr2: Box::new(exprs[2].clone()),
},
],
}),
};

*func = TableFunc::Repeat;
*exprs = vec![expr];
}
TableFunc::GenerateSeriesInt64 => {
// The arguments in `exprs` are `[start, stop, step]`.
// Per Brennan, the number of elements seems likely to be max (0, (stop - start + 1) / step)
let expr = MirScalarExpr::CallVariadic {
func: VariadicFunc::Greatest,
exprs: vec![
MirScalarExpr::literal_ok(
mz_repr::Datum::Int64(0),
mz_repr::ScalarType::Int64,
),
MirScalarExpr::CallBinary {
func: BinaryFunc::DivInt64,
expr1: Box::new(MirScalarExpr::CallBinary {
func: BinaryFunc::SubInt64,
expr1: Box::new(MirScalarExpr::CallBinary {
func: BinaryFunc::AddInt64,
expr1: Box::new(exprs[1].clone()),
expr2: Box::new(MirScalarExpr::literal_ok(
mz_repr::Datum::Int64(1),
mz_repr::ScalarType::Int64,
)),
}),
expr2: Box::new(exprs[0].clone()),
}),
expr2: Box::new(exprs[2].clone()),
},
],
};

*func = TableFunc::Repeat;
*exprs = vec![expr];
}
_ => {
// Do nothing for table functions we do not recognize.
// Doing nothing is a safe default.
}
}
}

columns_to_pushdown.retain(|c| *c < inner_arity);

reverse_permute(exprs.iter_mut(), columns_to_pushdown.iter());
Expand Down