Skip materializing splitless chunk children during split collection - #9271
Merged
Conversation
AdamGS
force-pushed
the
adamg/split-bench
branch
from
August 7, 2026 11:35
acf6eef to
e09ea5d
Compare
AdamGS
force-pushed
the
adamg/bench-optimization-experiment
branch
from
August 7, 2026 11:35
37a60b9 to
fb7d366
Compare
Merging this PR will improve performance by ×8.6
Performance Changes
Tip Curious why this is faster? Comment Comparing 
Footnotes
|
AdamGS
added
the
action/bench-sql
github-actions
Bot
removed
the
action/bench-sql
AdamGS
added
action/bench-random-access
github-actions
Bot
removed
action/bench-compress
Contributor
Polar Signals Profiling ResultsLatest Run
Powered by Polar Signals Cloud |
Contributor
Benchmarks: TPC-H SF=1 on NVME 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed / ns (1.003x ➖, 0↑ 0↓)
datafusion / parquet / ns (0.985x ➖, 2↑ 2↓)
duckdb / vortex-file-compressed / ns (0.958x ➖, 4↑ 0↓)
duckdb / parquet / ns (0.997x ➖, 0↑ 0↓)
File Size Changes (9 files changed, -43.9% overall, 0↑ 9↓)
Totals:
|
Contributor
Benchmarks: Clickbench Sorted on NVME 📖Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed / ns (0.890x ✅, 3↑ 0↓)
datafusion / parquet / ns (0.964x ➖, 2↑ 1↓)
duckdb / vortex-file-compressed / ns (0.989x ➖, 1↑ 1↓)
duckdb / parquet / ns (1.002x ➖, 0↑ 0↓)
File Size Changes (201 files changed, -42.8% overall, 51↑ 150↓)
Totals:
|
Contributor
Benchmarks: TPC-H SF=1 on S3 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed / ns (0.833x ➖, 4↑ 0↓)
datafusion / parquet / ns (1.064x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed / ns (1.038x ➖, 0↑ 3↓)
duckdb / parquet / ns (1.036x ➖, 0↑ 0↓)
|
Contributor
Benchmarks: FineWeb NVMe 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed / ns (0.993x ➖, 0↑ 0↓)
datafusion / parquet / ns (0.976x ➖, 1↑ 0↓)
duckdb / vortex-file-compressed / ns (0.986x ➖, 1↑ 0↓)
duckdb / parquet / ns (1.001x ➖, 0↑ 0↓)
File Size Changes (2 files changed, -46.3% overall, 0↑ 2↓)
Totals:
|
Contributor
Benchmarks: PolarSignals Profiling 📖Vortex (geomean): 1.005x ➖ datafusion / vortex-file-compressed / ns (1.005x ➖, 1↑ 1↓)
No file size changes detected. |
Contributor
Benchmarks: TPC-H SF=10 on NVME 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed / ns (0.998x ➖, 0↑ 0↓)
datafusion / parquet / ns (1.006x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed / ns (1.046x ➖, 0↑ 2↓)
duckdb / parquet / ns (0.987x ➖, 1↑ 0↓)
File Size Changes (9 files changed, -44.0% overall, 0↑ 9↓)
Totals:
|
Contributor
Benchmarks: Clickbench on NVME 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed / ns (0.986x ➖, 2↑ 1↓)
datafusion / parquet / ns (0.999x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed / ns (1.001x ➖, 1↑ 0↓)
duckdb / parquet / ns (1.011x ➖, 1↑ 3↓)
File Size Changes (101 files changed, -39.2% overall, 0↑ 101↓)
Totals:
|
Contributor
Benchmarks: Statistical and Population Genetics 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
duckdb / vortex-file-compressed / ns (0.970x ➖, 2↑ 1↓)
duckdb / parquet / ns (0.997x ➖, 0↑ 0↓)
File Size Changes (2 files changed, -32.3% overall, 0↑ 2↓)
Totals:
|
Contributor
Benchmarks: TPC-DS SF=1 on NVME 📖Verdict: No clear signal (low confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed / ns (1.010x ➖, 0↑ 0↓)
datafusion / parquet / ns (0.996x ➖, 1↑ 0↓)
duckdb / vortex-file-compressed / ns (1.019x ➖, 3↑ 8↓)
duckdb / parquet / ns (1.002x ➖, 2↑ 4↓)
File Size Changes (25 files changed, -43.5% overall, 0↑ 25↓)
Totals:
|
Contributor
Benchmarks: Random Access 📖Verdict: Likely improvement (low confidence) How to read Verdict and Engines
random-access / vortex-file-compressed / ns (0.848x ✅, 4↑ 0↓)
random-access / parquet / ns (1.006x ➖, 0↑ 0↓)
random-access / lance / ns (0.998x ➖, 0↑ 0↓)
|
Contributor
Benchmarks: FineWeb S3 📖Verdict: No clear signal (environment too noisy confidence) How to read Verdict and Engines
datafusion / vortex-file-compressed / ns (1.955x ❌, 0↑ 6↓)
datafusion / parquet / ns (0.991x ➖, 0↑ 0↓)
duckdb / vortex-file-compressed / ns (0.975x ➖, 0↑ 0↓)
duckdb / parquet / ns (1.275x ➖, 0↑ 1↓)
|
Benchmarks SplitBy::Layout over written files (struct -> chunked -> flat
per column) across column and chunk counts:
- cold: fresh reader tree per iteration, as the first scan over a file
- warm: reused reader tree
- cold_single_column: narrow field mask over a wide schema
- {cold,warm}_misaligned: per-column chunk granularities so no two
columns share interior chunk boundaries
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
Split collection previously built a full Layout + LayoutReader for every chunk of every column just so flat chunks could re-register the chunk-end boundary the parent already knows from its chunk offsets. - VTable::registers_interior_splits() (default true; flat returns false) lets a layout declare that its readers only ever push the end of the requested range, forwarded through LayoutVTablePlugin and DynLayout. - LayoutChildren::child_has_no_interior_splits(idx) answers that without materializing the child: owned children ask the vtable, viewed children resolve only the flatbuffer encoding tag against the registry, with a single-slot memo since siblings almost always share one encoding. - ChunkedReader caches a lazy per-chunk classification; when no chunk has interior splits it bulk-extends boundaries straight from chunk_offsets. - RowSplits drops consecutive identical ascending runs (columns with aligned chunk boundaries) and skips the final sort when a single run survives. StructReader pushes its end boundary after the field walk so the common case stays one ascending run. - LazyReaderChildren::new_uniform avoids a DType and name clone per chunk at reader construction. Split-collection bench medians (64 columns x 256 chunks): cold 916us -> 181us, warm 247us -> 13us; with fully misaligned columns cold 1223us -> 321us, warm 387us -> 146us. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Signed-off-by: Adam Gutglick <adam@spiraldb.com>
AdamGS
force-pushed
the
adamg/split-bench
branch
from
August 7, 2026 12:56
e09ea5d to
7274546
Compare
AdamGS
force-pushed
the
adamg/bench-optimization-experiment
branch
from
August 7, 2026 12:56
fb7d366 to
0ada06f
Compare
Contributor
Benchmarks: Compression 📖vortex / vortex-file-compressed / ns (1.002x ➖, 0↑ 1↓)
vortex / vortex-file-compressed / bytes (1.000x ➖, 0↑ 0↓)
vortex / vortex-file-compressed / ratio (1.001x ➖, 0↑ 0↓)
vortex / parquet / ns (1.001x ➖, 0↑ 0↓)
vortex / parquet / bytes (1.000x ➖, 0↑ 0↓)
|
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
Comment on lines
+103
to
+108
| pub struct RowSplits { | ||
| splits: Vec<u64>, | ||
| /// Start index of the run currently being appended. | ||
| run_start: usize, | ||
| /// Start index of the surviving run immediately before `run_start`. | ||
| prev_run_start: usize, |
Contributor
There was a problem hiding this comment.
I think you want a sorted; Vec<Vec>, Vec unsorted
then do n-way merge after?
Contributor
Author
There was a problem hiding this comment.
benchmark it and seems like the plan sort_unstable just beats it
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
AdamGS
force-pushed
the
adamg/bench-optimization-experiment
branch
from
August 7, 2026 15:15
acbf808 to
1d9a244
Compare
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
AdamGS
force-pushed
the
adamg/bench-optimization-experiment
branch
from
August 7, 2026 16:02
873aa08 to
98aa23a
Compare
Signed-off-by: Adam Gutglick <adam@spiraldb.com>
joseph-isaacs
approved these changes
Aug 7, 2026
joseph-isaacs
marked this pull request as ready for review
August 7, 2026 16:24
alexey-milovidov
added a commit
to ClickHouse/ClickHouse
that referenced
this pull request
Aug 21, 2026
0.84.0 carries vortex-data/vortex#9271, which stops split collection from building a `Layout` and a `LayoutReader` for every chunk of every column just to recover the chunk boundaries the parent already knows. That setup cost is what a short query on a wide file pays: on 0.83.0, `SELECT * ... LIMIT 1` over a 100-column file takes 0.31 s at 1 000 000 rows and 0.70 s at 4 000 000, against 0.15 s when a single column is selected, so it grows with chunks times columns rather than with the size of the result. Reported by @m7kss1, who measured 8.5 s for `SELECT * FROM hits LIMIT 1` on a file of 100 000 000 rows. The only API change that reaches this crate is that `ScanBuilder::with_projection` and `with_filter` now take a `BoundExpression`, so the projection and the pushed filter are bound to the file's type first. Binding type-checks the whole expression in one walk, which is where a column that is not in the file, or a comparison between two types that cannot be compared, is caught now. The vendored crates are in ClickHouse/rust_vendor#79. The lock file was updated with `cargo update -p vortex --precise 0.84.0` rather than regenerated, so nothing outside the `vortex` dependency tree moves: 37 `vortex-*` crates, `fastlanes` 0.5.2 -> 0.6.1, `fsst-rs` 0.5.11 -> 0.6.0, `onpair` 0.1.1 -> 0.2.0, `pastey` added, `core_detect` dropped.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changes are included in this PR?
Split collection previously built a full Layout + LayoutReader for every chunk of every column just so flat chunks could re-register the chunk-end boundary the parent already knows from its chunk offsets.
VTable::registers_interior_splits()(default true; flat returns false) lets a layout declare that its readers only ever push the end of the requested range, forwarded through LayoutVTablePlugin and DynLayout.LayoutChildren::child_has_no_interior_splits(idx)answers that without materializing the child: owned children ask the vtable, viewed children resolve only the flatbuffer encoding tag against the registry, with a single-slot memo since siblings almost always share one encoding.ChunkedReadercaches a lazy per-chunk classification; when no chunk has interior splits it bulk-extends boundaries straight from chunk_offsets.RowSplitsdrops consecutive identical ascending runs (columns with aligned chunk boundaries) and skips the final sort when a single run survives. StructReader pushes its end boundary after the field walk so the common case stays one ascending run.LazyReaderChildren::new_uniformavoids a DType and name clone per chunk at reader construction.Split-collection bench medians (64 columns x 256 chunks): cold 916us ->
181us, warm 247us -> 13us; with fully misaligned columns cold 1223us -> 321us, warm 387us -> 146us.