Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
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
14 changes: 14 additions & 0 deletions Cargo.lock

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

2 changes: 2 additions & 0 deletions Cargo.toml
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,7 @@ members = [
"encodings/bytebool",
"encodings/parquet-variant",
"encodings/onpair",
"encodings/dense-union",
# Benchmarks
"benchmarks/lance-bench",
"benchmarks/compress-bench",
Expand Down Expand Up @@ -304,6 +305,7 @@ vortex-compute = { version = "0.1.0", path = "./vortex-compute", default-feature
vortex-datafusion = { version = "0.1.0", path = "./vortex-datafusion", default-features = false }
vortex-datetime-parts = { version = "0.1.0", path = "./encodings/datetime-parts", default-features = false }
vortex-decimal-byte-parts = { version = "0.1.0", path = "encodings/decimal-byte-parts", default-features = false }
vortex-dense-union = { version = "0.1.0", path = "./encodings/dense-union", default-features = false }
vortex-edition = { version = "0.1.0", path = "./vortex-edition", default-features = false }
vortex-error = { version = "0.1.0", path = "./vortex-error", default-features = false }
vortex-fastlanes = { version = "0.1.0", path = "./encodings/fastlanes", default-features = false }
Expand Down
33 changes: 33 additions & 0 deletions encodings/dense-union/Cargo.toml
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@
[package]
name = "vortex-dense-union"
authors = { workspace = true }
categories = { workspace = true }
description = "Dense union encoding for Vortex arrays"
edition = { workspace = true }
homepage = { workspace = true }
include = { workspace = true }
keywords = { workspace = true }
license = { workspace = true }
readme = "README.md"
repository = { workspace = true }
rust-version = { workspace = true }
version = { workspace = true }

[dependencies]
prost = { workspace = true }
vortex-array = { workspace = true }
vortex-error = { workspace = true }
vortex-mask = { workspace = true }
vortex-session = { workspace = true }

[dev-dependencies]
divan = { workspace = true }
vortex-array = { workspace = true, features = ["_test-harness"] }
vortex-buffer = { workspace = true }

[lints]
workspace = true

[[bench]]
name = "take"
harness = false
3 changes: 3 additions & 0 deletions encodings/dense-union/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,3 @@
# Vortex Dense Union

An external dense physical encoding for Vortex's logical `DType::Union`.
118 changes: 118 additions & 0 deletions encodings/dense-union/benches/take.rs

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

This benchmark never runs in CI. .github/workflows/codspeed.yml drives benchmarks off an explicit package matrix (9 shards, four of them "Encodings 1" through "Encodings 4"), and vortex-dense-union is in none of them — so the numbers in the PR description cannot regress a build, and the 2.1x gap has nothing watching it.

Adding the package to one of the encodings shards is a one-line change (shard 4 and shard 7 are the shortest lists). No features entry is needed: the _test-harness feature is already pulled in through this crate's dev-dependencies.

Given the PR ships a known 2.1x regression versus sparse and the canonicalize hot loop has obvious headroom, this is the check that would tell you whether the follow-up optimizations actually landed.


Generated by Claude Code

Original file line number Diff line number Diff line change
@@ -0,0 +1,118 @@
// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: Copyright the Vortex contributors

#![expect(clippy::unwrap_used)]
#![expect(clippy::cast_possible_truncation)]

use std::sync::LazyLock;

use divan::Bencher;
use vortex_array::ArrayRef;
use vortex_array::IntoArray;
use vortex_array::RecursiveCanonical;
use vortex_array::VortexSessionExecute;
use vortex_array::array_session;
use vortex_array::arrays::PrimitiveArray;
use vortex_array::arrays::UnionArray;
use vortex_array::dtype::DType;
use vortex_array::dtype::FieldNames;
use vortex_array::dtype::Nullability;
use vortex_array::dtype::PType;
use vortex_array::dtype::UnionVariants;
use vortex_dense_union::DenseUnion;
use vortex_dense_union::initialize;
use vortex_session::VortexSession;

const LEN: usize = 65_536;
const N_VARIANTS: usize = 28;
const TAKE_LEN: usize = 4_096;

fn main() {
LazyLock::force(&SESSION);
divan::main();
}

static SESSION: LazyLock<VortexSession> = LazyLock::new(|| {
let session = array_session();
initialize(&session);
session
});

fn variants() -> UnionVariants {
let names = FieldNames::from_iter((0..N_VARIANTS).map(|index| format!("variant_{index}")));
let dtypes = vec![DType::Primitive(PType::I32, Nullability::NonNullable); N_VARIANTS];
let type_ids = (1..=N_VARIANTS).map(|type_id| type_id as u8).collect();
UnionVariants::try_new(names, dtypes, type_ids).unwrap()
}

fn selectors() -> (ArrayRef, ArrayRef, Vec<usize>) {
let mut child_lengths = vec![0usize; N_VARIANTS];
let mut type_ids = Vec::with_capacity(LEN);
let mut offsets = Vec::with_capacity(LEN);
for row in 0..LEN {
let child_index = row % N_VARIANTS;
type_ids.push((child_index + 1) as u8);
offsets.push(child_lengths[child_index] as i32);
child_lengths[child_index] += 1;
}
(
PrimitiveArray::from_iter(type_ids).into_array(),
PrimitiveArray::from_iter(offsets).into_array(),
child_lengths,
)
}

fn dense_union() -> ArrayRef {
let (type_ids, offsets, child_lengths) = selectors();
let children = child_lengths
.into_iter()
.map(|len| PrimitiveArray::from_iter(0..len as i32).into_array())
.collect::<Vec<_>>();
DenseUnion::try_new(type_ids, offsets, variants(), children)
.unwrap()
.into_array()
}

fn sparse_union() -> ArrayRef {
let (type_ids, ..) = selectors();
let children = (0..N_VARIANTS)
.map(|child_index| {
PrimitiveArray::from_iter((0..LEN).map(move |row| {
if row % N_VARIANTS == child_index {
(row / N_VARIANTS) as i32
} else {
0
}
}))
.into_array()
})
.collect::<Vec<_>>();
UnionArray::try_new(type_ids, variants(), children)
.unwrap()
.into_array()
}

fn indices() -> ArrayRef {
PrimitiveArray::from_iter((0..TAKE_LEN).rev().map(|index| index as u32)).into_array()
}

fn bench_take(bencher: Bencher, array: ArrayRef, indices: ArrayRef) {
bencher
.with_inputs(|| (&array, &indices, SESSION.create_execution_ctx()))
.bench_refs(|(array, indices, ctx)| {
array
.take((*indices).clone())
.unwrap()
.execute::<RecursiveCanonical>(ctx)
});
}

#[divan::bench]
fn dense_take(bencher: Bencher) {
bench_take(bencher, dense_union(), indices());
}

#[divan::bench]
fn sparse_take(bencher: Bencher) {
bench_take(bencher, sparse_union(), indices());
}
Comment on lines +111 to +118

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

You need to execute the array here otherwise there is no work that is done

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

Ah, that's correct, for now it is lazy-take. Fixed.

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

im interested in what the new benchmark results are?

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

Already showed in the PR content. DenseUnion is 2.10-2.12x slower.

Loading
Loading